⚡Megatron LLM Hub
全部
Website

rss-website

头条2

Brotli 解码缺陷可致 RCE,TLS/DNS 缺陷可劫持流量,各 LTS 线均受影响。

On October 21, 2026, Node.js announced the availability of fixes for multiple vulnerabilities across the following supported release lines: 22.x, 24.x, and 26.x. These vulnerabilities impact critical operations such as the Brotli decoder, DNS resolution, and TLS certificate validation. The vulnerabilities pose risks for remote code execution and traffic interception. For more information, see Node.js Security Release . The following vulnerabilities have a Security Impact Rating (SIR) of High and are addressed in this advisory: CVE-2026-111497 - Brotli decoder out-of-bounds write; potential remote code execution CVE-2026-105482 - DNS resolution truncation involving embedded NUL characters CVE-2026-105483 - TLS hostname validation bypass using fullwidth IP digits CVE-2026-106069 - TLS hostname validation bypass with noncanonical IP addresses Important: If a product is not listed within this advisory, Cisco's review indicates that the product is not affected by this vulnerability. This advisory will be updated as additional information becomes available. This advisory is available at the following link: https://sec.cloudapps.cisco.com/security/center/content/CiscoSecurityAdvisory/cisco-sa-oct2026-node-js-Jj3SQ19u <br/>Security Impact Rating: High

rcetlsnode.jscve @cisco-advisory 原文 ↗
必看3

邮件服务器多暴露于公网且此漏洞无需凭据即可触发,攻击复杂度高但影响面广。

MailEnable Auth Cleartext Storage Information Disclosure Vulnerability October 27th, 2026 ZDI-26-810 ZDI-CAN-34478 CVE ID CVE-2026-114551 CVSS Score 5.9 AV:N/AC:H/PR:N/UI:N/S:U/C:H/I:N/A:N Affected Vendors MailEnable Affected Products MailEnable Vulnerability Details This vulnerability allows remote attackers to disclose sensitive information on affected installations of MailEnable. Authentication is not required to exploit this vulnerability. The specific flaw exists within the handling of mailbox credentials. The issue results from storing sensitive information in plaintext. An attacker can leverage this in conjunction with other vulnerabilities to disclose stored credentials, leading to further compromise. Additional Details Fixed in version 10.60 https://www.mailenable.com/Premium-ReleaseNotes.txt Disclosure Timeline 2026-09-18 - Vulnerability reported to vendor 2026-10-27 - Coordinated public release of advisory 2026-10-27 - Advisory Updated Credit Nicolas Chatelain Back to Advisories

cve信息泄露邮件服务器 @zdi-advisories 原文 ↗

浏览器内类型混淆是稳定 RCE 路径,需确认终端浏览器版本已更新。

# CVE-2026-62765 - Microsoft Edge (Chromium-based) Remote Code Execution Vulnerability Access of resource using incompatible type ('type confusion') in Microsoft Edge (Chromium-based) allows an unauthorized attacker to execute code over a network. CWE: CWE-843: Access of Resource Using Incompatible Type ('Type Confusion')

rce浏览器type_confusioncve @msrc 原文 ↗

路径穿越成因与利用分析较具体,自托管财务应用常直接暴露在公网,凭据与账本文件都在读取范围内。

Introduction In modern web applications, handling file paths securely is a critical requirement. When input fields accept file paths or file names directly without adequate validation, software can become susceptible to Directory Traversal (Path Traversal) flaws. CVE-2026-116067 represents a high-severity Path Traversal vulnerability discovered in Paisa—a popular open-source, self-hosted personal finance tracking application designed to help users manage their ledgers, budgets, and investments. Self-hosted financial applications often process sensitive personal data, financial records, and system configuration credentials. When security controls around filesystem access fail, the exposure of these underlying operating system files creates a massive privacy and operational risk. In multi-tenant environments or instances exposed directly to local networks and the public internet, unauthenticated flaws bypass traditional perimeter defenses entirely. This article provides a comprehensive technical teardown of CVE-2026-116067, detailing its underlying architecture, exploitation vectors, impact, and concrete remediation steps. Learning Objectives By reading this analysis, security engineers, software developers, and system administrators will learn: Understand Path Traversal Mechanics: How path traversal characters ( ../ or ..\ ) bypass filesystem sandboxes. Identify Insecure File API Endpoint Handlers: Understand how improperly sanitized HTTP input parameters expose internal files to remote users. Analyze CVE-2026-116067: Examine the root cause affecting Paisa versions prior to 0.7.8 . Implement Proper Sanitization: Recognize code-level defenses to safely resolve relative path references in file-handling applications. What is Paisa < 0.7.8 – Unauthenticated Path Traversal (CVE-2026-116067) CVE-2026-116067 is an unauthenticated path traversal vulnerability belonging to the CWE-22 weakness classification. The flaw affects all releases of the Paisa personal finance tracking platform prior to version 0.7.8 . It carries a CVSS 3.1 base score of 8.8 (High) , reflecting severe risk to data confidentiality and application integrity. The core of the issue lies in how Paisa handles remote file requests sent to its HTTP server components. Because Paisa is designed for simple, out-of-the-box local or self-hosted deployment, many setups run without mandatory authentication proxies or built-in login barriers. Consequently, any network adversary capable of reaching the application’s HTTP port can craft malicious payloads that reference internal directories outside the web root directory. When successfully exploited, an unauthenticated remote attacker gains arbitrary read access to sensitive system files, application secrets, and configuration documents. Furthermore, because specific affected endpoints govern backup management, attackers can also trigger unauthorized file deletions across the target filesystem, restricted only by the permissions of the operating system user running the Paisa process. Key aspects of this vulnerability include: CVSS Score & Severity: Rated 8.8 High due to remote unauthenticated access and combined read/delete capabilities. Vulnerability Type: Classified under CWE-22 (Improper Limitation of a Pathname to a Restricted Directory) . Authentication Requirement: Zero authentication required to deliver the malicious payloads over HTTP. Primary Exploitation Risks: Unauthorized exfiltration of sensitive files (e.g., /file/path , private keys) and arbitrary deletion of backup files. +-------------------+ Malicious Payload +----------------------+ | | GET /api/file?name=../../../../file/path | | | Unauthenticated |------------------------------------------->| Paisa Server | | Remote Attacker | | (Version < 0.7.8) | | |<= = = = = = = = = = = = = = = = = = = = = =| | +-------------------+ Exfiltrated Host System File +----------------------+ | Unsanitized Access v +----------------------+ | Host Filesystem | | / file/path | +----------------------+ Technical Detail: How the Vulnerability Works The technical root cause stems from missing path canonicalization and input sanitization inside Paisa’s backend REST API handlers. When processing incoming client requests that reference file storage or ledger backups, the backend concatenates user-controlled string inputs directly into target filesystem operations without scrubbing parent directory markers ( ../ ). Specifically, four primary API endpoints fail to strip path traversal sequences prior to executing file reads or file deletions. By appending sequential relative directory traversals, an attacker breaks out of the intended working directory ( /app/data or /ledger ) and traverses up the directory tree toward the host system root directory. Once the application resolves the requested relative path, it opens or deletes the file under the context of the running Paisa daemon. If the process is run under elevated privileges or a standard user account with broad read access, the entire accessible filesystem becomes vulnerable to data exposure or targeted file destruction. Specific vulnerable endpoints and operational traits: GetFile Handler: Fails to sanitize file parameters, allowing remote arbitrary file retrieval across the host filesystem. GetSheet Handler: Processes sheet path inputs directly, enabling unauthorized viewing of internal system documents. DeleteBackups Handler: Accepts traversal sequences inside backup deletion requests, exposing host files to deletion. DeleteSheetBackups Handler: Lacks filename verification, allowing remote removal of targeted files outside the backup directory. The terminal output below illustrates an unauthenticated attacker retrieving / file/path via a crafted HTTP GET request: Affected Software & Plugins Paisa is a standalone personal finance manager built to give users complete control over their financial records using plain-text accounting principles. Because the application includes an embedded web server for interface rendering and backup management, any vulnerable deployment exposed to a network is directly susceptible to CVE-2026-116067 without requiring secondary plugins or extensions. The vulnerability impacts all instances running versions prior to the official patch release. Administrators should review their deployment versions against the following summary metrics: Target Application: Paisa (Personal Finance Management Software) Vulnerable Versions: All releases prior to 0.7.8 ( < 0.7.8 ) Fixed Version: Release 0.7.8 (Git commit 035949351f7c0d0f90ecf767aa1e28b4d5a0ae31 ) Attack Prerequisites: Network line-of-sight to the Paisa HTTP server port; no active authentication required. Conclusion CVE-2026-116067 serves as a stark reminder of the critical need for strict input validation when software interacts directly with the host filesystem. Unauthenticated path traversal vulnerabilities compromise the core security boundary of self-hosted applications, exposing sensitive internal state files and user data to remote exploit automation. Software developers must ensure that user-supplied path inputs are consistently sanitized using robust path resolution methods, such as validating canonical file paths against strict base-directory prefixes. Relying solely on client-side controls or basic string concatenation creates severe security blind spots that attackers easily exploit. Organizations and individuals self-hosting Paisa instances should immediately update their installations to version 0.7.8 or later . If immediate upgrading is not feasible, administrators should isolate Paisa behind an authenticating reverse proxy or restrict access strictly to trusted local loopback interfaces ( 127.0.0.1 ). Mastering the Linux Terminal: From Basics to Advanced Commands CVE-2026-29059: Unauthenticated Path Traversal in Windmill and Nextcloud Flow Vulnerability Analysis CVE-2026-87902 WordPress Core – PHP Template Path Traversal Severe Threat in Adobe ColdFusion: CVE-2026-48282 RDS Arbitrary File Write Vulnerability Vulnerability Analysis CVE-2026-55224 MineAdmin < 3.2.0-alpha.2 – Plugin Path Traversal to RCE The post Vulnerability Analysis: CVE-2026-116067 Paisa < 0.7.8 – Unauthenticated Path Traversal appeared first on DenizHalil - Professional Cybersecurity Consulting and Penetration Testing .

路径穿越cveweb @denizhalil-blogs 原文 ↗
推荐5

认证后 XXE 可与凭据泄露链式利用,是邮件网关类产品的典型攻击面。

MailEnable parseObjectResponse XML External Entity Processing Information Disclosure Vulnerability October 27th, 2026 ZDI-26-811 ZDI-CAN-29137 CVE ID CVE-2026-114552 CVSS Score 6.5 AV:N/AC:L/PR:L/UI:N/S:U/C:H/I:N/A:N Affected Vendors MailEnable Affected Products MailEnable Vulnerability Details This vulnerability allows remote attackers to disclose sensitive information on affected installations of MailEnable. Authentication is required to exploit this vulnerability. The specific flaw exists within the implementation of the parseObjectResponse method. The issue results from the lack of proper restriction of XML External Entity (XXE) references. An attacker can leverage this vulnerability to disclose information in the context of the service account. Additional Details Fixed in version 10.60 https://www.mailenable.com/Premium-ReleaseNotes.txt Disclosure Timeline 2026-09-18 - Vulnerability reported to vendor 2026-10-27 - Coordinated public release of advisory 2026-10-27 - Advisory Updated Credit Nicolas Chatelain Back to Advisories

xxecve邮件服务器 @zdi-advisories 原文 ↗

给出 agent 自主性与人类监督边界的实测设计,直接关系 AI 安全治理。

Chapter 01 : Study design Amaç Herdağdelen, Hailey Joren, Alex Tamkin When experts hand their work to an AI agent, the usual assumption is that they still make the decisions that matter. The agent writes the code, runs the numbers, and offers advice, but the expert decides what to measure, how to interpret ambiguous evidence, and what the next steps should be. Often there is more than one reasonable answer to these questions, and the decision will come down to the experience and values of whoever makes the call. We wanted to look at whether experts really do make those important decisions, or whether in practice they defer to their AI agents. In a study we report in a new paper , we asked 50 professional analysts to do a realistic statistical analysis with Claude Code. They all had experience in causal inference—that is, working out from data whether one thing causes another. Their task was to advise a farming cooperative on whether an advisory service it was running was effective and should be expanded. Along the way, the analysts had to make four choices between two reasonable options, such as whether to measure productivity as crop yield per hectare or as yield per unit of operating cost, and then make a final decision. To see whether they were making these calls themselves, we set up each analyst’s agent to favor one option on each decision, which the analysts weren’t told about. Which way the agents would “lean” would be decided by coin flip. The agent was told to make the best honest case for its lean, to change its mind if the evidence demanded it, and never to mislead. Across the four analytical choices, analysts went with their agent’s lean 86% of the time. If the agent’s input made no difference, we’d expect about 50%. When analysts had already expressed their own preference before the agent weighed in, they switched away from it 60% of the time if the agent disagreed, compared to 6% if it agreed. Afterward, when asked to explain their choices, 77% of analysts used the agents’ own arguments—sometimes word for word—but only 15% credited the agent as the source of those arguments. Analysts who followed the agents’ recommendations closely were about as confident in their final recommendation as the rest, and roughly as likely to call it the product of their own judgment. Participants work through an analysis task with five open-ended decisions. For each participant and decision, a coin flip sets which option their AI leans toward, without telling the participant. We compare how often participants choose option A under each lean. We don’t know how much this tells us about the real world. Real agents will sometimes lean one way or another—with no lean assigned, Claude Opus 4.8 made the same five choices in 112 out of 123 runs—but in this case we strengthened their preferences on purpose to measure their effect. When agents aren’t given clear preferences and instructions to voice them, the effect might be smaller. This isn’t a simple cautionary tale against working with AI. Analysts who used the agent finished the task faster and made far fewer serious errors than those who chose to work without it, although the latter group was small and less experienced, so we are wary of reading too much into it. But as AI takes over more of the hands-on work, decisions like these are increasingly where human judgment and values will affect the result. If experts defer to agents without realizing it, the agents’ preferences—which could be arbitrary or misaligned—could shape important decisions about policy proposals or research directions without anyone having chosen those preferences. This is particularly important for research into keeping AI safe. Researchers have begun using agents to run and analyze experiments and a human reviewer is only an independent check if their judgment is their own. Study design We recruited 61 analysts for a randomized controlled trial. All came from a pre-screened pool of professionals who had passed a causal-inference knowledge check, used the programming languages Python or R regularly, and had hands-on experience with observational causal analysis, which means studying cause and effect using real-world data rather than a controlled experiment. The task was set at a fictional organization, the Greenfields Farming Cooperative, which runs an (equally fictitious) voluntary, paid advisory program to help its member farms improve their crop yields. About 40% of members have enrolled. The co-op’s board must decide whether to subsidize and expand the program, or discontinue it. The board will expand the program only if it improves farms’ productivity by a set margin, roughly 3–4%, but it doesn’t say exactly how productivity should be measured. We deliberately left four analytical decisions open, each with two reasonable options: Outcome : measure productivity as crop yield per hectare, or as yield per unit of operating cost? Comparison method : look at how enrolled farms’ yields changed after joining compared with how other farms’ yields changed over the same period (a “difference-in-differences” design), or compare yields directly while adjusting statistically for differences between the farms? Which farms the assessment covers : measure the effect only in the enrolled farms, or across the cooperative’s entire membership? Weighting : count every farm equally, or give larger producers more weight? These four choices led into a final recommendation: whether to roll out the program to everyone, or to discontinue it. Analysts and agents could also propose a one-year extension. Each of the analytical decisions could move the estimated productivity gain by several percentage points, which is important when the board’s threshold is 3–4%. For each of the five decisions presented to each analyst, the agent’s lean was set by a coin flip. We gave the agent its leans in hidden instructions (a system prompt) that told it to hold each lean as its own view, to raise it when the decision came up, and to make the strongest case for it that the data allowed. The same instructions told it to change its mind if the evidence turned decisively against it, and never to misrepresent results, hide data, or make an argument the numbers didn’t support. Out of 61 participants, 11 chose not to use the agent and are left out of the analysis unless noted. Among the remaining 50, 65% held a master’s or doctorate or had six or more years of experience, and 90% rated themselves proficient or expert in observational causal inference (of the 48 who answered the survey). The median time on the task was 65 minutes, and half of the analysts took between 50 and 79 minutes. They worked in a web-based version of the VS Code editor with Claude Code built in, running Claude Opus 4.8 at High effort. When they finished, analysts submitted their four analytical choices and their final recommendation alongside an estimate of the program’s effect on productivity, with a confidence interval. They then completed a post-task survey. A separate AI model read each submission and its code to record which option the analyst had chosen on each decision. Where its reading disagreed with the analyst’s own survey answer, we made the final call. Results Analysts mostly went with their agents’ lean. On each of the four analytical decisions, analysts’ choices matched their agents’ lean the large majority of the time: 92% for the outcome measure, 82% for the comparison method, 81% for which farms the estimate covered, and 90% for weighting. Across all four combined, the result was 86%. Figure 2 shows how much the agents’ lean shifted analysts’ choices on each decision, including the final recommendation. Agents’ lean had a weaker effect on the final recommendation, but its earlier input still shaped the result. When the agent leaned toward rolling out the program, 54% of analysts recommended it, compared with 36% when it leaned the other way. (A one-year extension counts as not rolling out.) We think the effect was weaker because, once the analysts had their estimate, the recommendation largely followed from whether it cleared the board’s 3–4% threshold, which left less room for judgment. But by then, an agent’s lean on the analytical decisions had already shaped the estimate itself. Using only those four leans, a statistical model correctly predicted 82% of the analysts’ recommendations, compared with 61% for always guessing the most common recommendation. (This excludes analysts who proposed an extension.) So we could largely predict what an analyst would recommend from the coin flips alone, before they had even started the task. The share of participants choosing option A on each decision, when their AI leaned A (green) and when it leaned B (violet). The four analytical decisions moved 62 to 82 percentage points and the final recommendation 17. Analysts who had already picked an option often switched when their agent disagreed. In 38 cases, an analyst had expressed a preference on one of the analytical decisions before the agent expressed its lean. When the lean went against their choice, they switched 60% of the time, compared with 6% when it agreed with their choice. That’s a gap of 54 percentage points, but because this rests on only 38 cases, the estimate is imprecise, and the 95% confidence interval runs from 24 to 82 percentage points. We also can’t be sure that the difference is caused by the AI’s influence: the analysts who expressed a preference may have been systematically different in some way from those who didn’t. Experience and expertise did not predict who deferred. In an exploratory analysis, we checked whether any of 10 characteristics predicted how often analysts’ decisions followed their agents’ lean: years of experience, self-rated expertise in causal inference, education, score on a causal-inference quiz, trust in AI, how often they used AI, “need for cognition” (a measure of how much people enjoy effortful thinking), open-mindedness, gender, and age. Only age showed a statistically significant link, with older analysts following the lean more often. But we don’t read very much into that: if you test many different things, you’re more likely to get one positive result just by chance, and when we adjusted for that the result disappeared. Because most analysts followed their agent on most decisions, there was little variation to explain anyway, so even real differences would have been hard to detect. Analysts who followed their agent’s lean still called the recommendation their own. On the post-task survey, 52% of analysts rated the final recommendation as more their own than their agent’s. The rest said it was more the agent’s recommendation. But both groups were roughly equally influenced by their agents: the former’s choices shifted about 65 percentage points toward their agents’ lean, on average, compared to 79 percentage points for the latter, a difference within the margin of error. Analysts’ confidence in their own recommendation showed a similar pattern. They were asked to rate their confidence from 0 (“it’s a coin flip”) to 100 (“I’d stake my professional reputation on it”). Those who followed the lean on all four analytical decisions averaged 71, compared to 73 for the rest. On the post-task survey (n = 48), 52% called the recommendation more their own than the agent’s and 35% mostly or entirely their own — even though the agent’s leans had shifted their choices. Analysts explained their choices in their agent’s words. After the task, the survey asked about each decision: “Briefly, why that choice rather than the alternative? If it wasn’t an active choice, just say so.” Analysts couldn’t look back at their conversations with the agent while answering. Even so, 77% repeated their agent’s argument in at least one answer, sometimes word for word, and only 15% credited the agent anywhere. When explicitly asked what they thought the study was about, six of the 48 correctly guessed that it was about the AI’s influence on their choices, but none suspected that the AI’s own preferences had been manipulated. Agent statements during the task (orange) and the same participants’ answers afterward to “why that choice rather than the alternative?” (blue) share phrasing, sometimes verbatim. 77% of participants repeated their agent’s argument in at least one answer, 15% credited the agent. Participants could not go back to review the agent transcripts when writing these answers. Analysts who worked without AI made more serious errors and took longer. Of the 11 participants who chose not to use the agent, six (55%) made serious methodological errors that invalidated their results, compared with three of the 50 (6%) who did use it. Two AI models flagged the errors, and one of the study’s authors checked each one without knowing which group the analyst was in. The analysts who worked without an agent also took longer, with a median time of 90 minutes, compared to 65 for the others. The two groups were not directly comparable, so we can’t be sure whether the different outcomes were caused by the decision to use or not use AI. Those who worked without AI were less experienced on average: fewer held a master’s degree or a doctorate (four out of 11, compared to 24 out of 46 who used the agent), more had two years or less in analysis roles (six out of 11, compared to 15 out of 48), and fewer rated themselves proficient or expert in causal inference (eight of 11, compared to 43 out of 48). Our study’s limitations We deliberately strengthened the agents’ lean. Our instructions made the agent hold and voice a preference reliably, so that we could measure its effect. The rates we report show what happens when an agent leans one way and says so, but how much that applies to real-world work will depend on how strong the expert’s own views are, how clearly the decisions involved are defined, and how the model behaves, including whether it has a preference and whether it volunteers it. We studied one task, in one field, with one model. The task took about an hour, it was specifically about farming, and the agent ran on Claude Opus 4.8. The analysts were experts in causal inference but new to farming. In real work, they would know the field better, and probably have stronger views of their own. They could also interview relevant people, talk to the board, and in other ways narrow down the reasonable options, which might reduce the agent’s influence. Our estimates are averages across analysts. Because the coin flips differed between analysts, comparing analysts whose agents leaned in opposite ways tells us the lean’s average effect. We didn’t record each participant’s own view before the agent voiced its lean, so we can’t say how much any single participant was influenced. We recorded which option analysts chose, but not how strongly they preferred it. One possible explanation for deferring is that the expert didn’t have strong opinions of their own. We measured only whether analysts’ choices shifted, not how confident they were in their choice. What this means One reason analysts deferred so readily may be that experts don’t always arrive with their views fully formed. They often work out, or even discover , what they think while weighing the options. The agents in our study were told not to mislead, and they were often open about how the results would change under each option. But they rarely made a strong case for why an expert might prefer the other option, which might have helped analysts make a better-informed choice. With only one side argued well, the analysts could adopt the agent’s assumptions without working out their own. For individual experts, this could mean losing agency over their own work, a form of what researchers call disempowerment . The work still goes out under the expert’s name, but the decisions that most depend on their judgment are in reality settled by the agent’s prior beliefs. The stakes extend beyond individuals, though, because so much of science and policy rests on judgment calls like these. When separate research teams reach the same result, we trust it more, because we assume each team reached it independently. If those teams all defer to models with the same preferences, their agreement may no longer be independent . An apparent consensus might reflect one model’s preferences, and if those preferences are arbitrary or misaligned, the same bias would shape many results at once. The same problem applies to AI safety research. Researchers have begun to use AI agents to run and analyze experiments, in part because there is more work than human staff can keep up with. Keeping a human in the loop assumes that the human’s judgment provides an independent check. But in our study, analysts signed off on work in which judgment calls had largely been settled by the agent’s randomly assigned preferences. Our full paper also shows that agents without an assigned lean have weaker preferences of their own, and sometimes push back against users. Deferral of this kind is hard to detect. Checking AI-assisted work for errors, or against a reference answer, assumes there is a correct answer. But where there isn’t one, as in judgment calls, the work can look sound whichever option is chosen. Asking experts afterward may not help either, because they may not notice the agent’s influence. This isn’t an argument for using AI less. Used well, AI could expand people’s agency rather than eroding it. We suggest three areas for future study: Measuring deferral. We could measure deferral in our study because we randomized the agent’s lean, which is rarely possible in real work, and our results suggest that people may not be able to report it accurately themselves. We need ways to measure it in everyday settings, and recent work that estimates how someone would have carried out a task without AI is a good start. We also need a better picture of models’ default preferences and how they express them, so that people can engage with models in a more informed way. Designing agents differently. Agents could make the best case for each option, or ask experts for their own view before offering one. But this could quickly become overwhelming for users. A single analysis can involve many open decisions, and much of the value of AI is that people don’t have to think through each one. Agents would need to raise only the decisions that matter most, but picking those is itself a judgment call. How a model speaks probably matters too, since confidence and fluency can sway people. A model that sounds very sure of itself might tilt people more than is warranted, but a mealy-mouthed model will just irritate the people who try to use it. Changing institutions. Institutions often treat an expert’s sign-off as evidence of independent judgment. As AI does more of the work, that assumption becomes shakier. Organizations will need to decide deliberately which decisions need a person’s own informed judgment, and give people the time and resources to form one. Judgment also depends on skill, and people who stop making these decisions may lose the ability to make them. As models become more capable, institutions will need an explicit strategy for establishing the role of humans and the development of their skills . Even with extremely capable AI systems, institutions may want to protect human skills and oversight for reasons of safety or resilience, or because they value them for their own sake. There may be no single right division of labor between people and AI. Some decisions may be fine to hand over, and others worth keeping. With the right tools, that split can be something we choose, instead of something we drift into. Authors & acknowledgements Acknowledgements This project was led by Amaç Herdağdelen, Hailey Joren, and Alex Tamkin. Editorial support for this blog post was provided by Tom Chivers and Arjun Panickssery. We would like to thank Joe Benton, Mojmir Stehlik, and Avery Griffin for making this project possible through the Anthropic Fellows Program. We would also like to thank Matthew Jörke, Onur Altındağ, and Vishakh Padmakumar for experimental design feedback. Citation Copy @misc{herdagdelen2026judgmentcalls, author = {Herdağdelen, Amaç and Joren, Hailey and Tamkin, Alex}, title = {How AI agents shape experts’ judgment calls}, year = {2026}, howpublished = {Anthropic Research}, url = {https://www.anthropic.com/research/agents-judgment-calls} }

ai_agentllm报告 @anthropic-research 原文 ↗

对 Themida 等加壳或加密 global-metadata 的 Unity 样本逆向有可直接复用的手法。

Download v1 https://www.unknowncheats.me/forum/d...=file&id=58964 This is a runtime version of my IL2CPP dumper. Instead of reading the metadata straight from disk, it runs inside the game and dumps the metadata that's already loaded in memory. I mainly made it for games where global-metadata.dat is encrypted, packed or otherwise protected. That includes common packers/protectors like Themida, and similar setups where the on-disk metadata is scrambled, encrypted, or not readable as a normal global-metadata.dat. Offline dumpers fail there because they try to parse the protected file. This one reads from the live process instead. Depending on how the game handles its metadata, IL2CPP may have already unpacked or decrypted what it needs into memory, so the dump can come from the runtime data rather than the protected disk copy. It won't work with games protected by anti-cheats like EAC, so don't try injecting it into those. Features - Dumps loaded assemblies, classes, fields and methods - Includes parent classes, interfaces, nested types and enums - Generates C++, C#, Rust, Python and JSON output - Includes an index and per-assembly C# files - Exports field offsets and method RVAs using names like _Offset and _RVA - Dumps metadata from memory to handle certain encrypted or packed metadata setups - Tries to handle renamed modules and stripped/missing IL2CPP exports Output The dump is saved to Desktop\GameDump, with C:\GameDump used as a fallback. Files include: - GameDump.hpp - GameDump.cs - GameDump.rs - GameDump.py - GameDump.json - Index.json - images\ containing per-assembly .cs files How to use 1. Start the game and wait for IL2CPP to initialise. The main menu is usually enough. 2. Inject dumper.dll into the process. 3. A console titled IL2CPP Dumper should open and show the progress. 4. Once it finishes, check the output folder. The dumper also tries to detect renamed modules, such as UserAssembly.dll, if the game doesn't use GameAssembly.dll. There's also a fallback scan for when the usual exports aren't available. Notes - x64 only - This release contains the runtime DLL only. - If the metadata is unprotected and readable on disk, a normal offline dumper is probably the easier option. - Intended for protected metadata setups; compatibility depends on how the game handles its metadata. If you run into any issues or find a game it doesn't handle properly, post the details here.

逆向工具脱壳 @unknowncheats-ac 原文 ↗

绕过类缺陷常被用于钓鱼与恶意投递链,需确认缓解策略仍生效。

# CVE-2026-96196 - Microsoft Edge (Chromium-based) Security Feature Bypass Vulnerability Incomplete list of disallowed inputs in Microsoft Edge (Chromium-based) allows an unauthorized attacker to bypass a security feature over a network. CWE: CWE-184: Incomplete List of Disallowed Inputs

绕过浏览器cve @msrc 原文 ↗

UI 伪造常用于诱导授权与凭据窃取,终端浏览器版本需核对。

# CVE-2026-20942 - Microsoft Edge (Chromium-based) Spoofing Vulnerability Improper restriction of rendered ui layers or frames in Microsoft Edge (Chromium-based) allows an unauthorized attacker to perform spoofing over a network. CWE: CWE-1021: Improper Restriction of Rendered UI Layers or Frames

浏览器ui伪造cve @msrc 原文 ↗
更多13

模块镂空是驱动级规避的通用手法,可用于理解反作弊与 EDR 的检测点。

I'm looking to make a cheat for Wardogs and potentially OW2, figured I'd start with a good driver setup. Looking to bypass Elytra and Warden, shouldn't be too hard. I've (GLM 5.3) modified KDMapper to module stomp instead of allocating random mem to store the driver. In my offline VM it places the driver into mrxsmb20.sys and all is well and good. Obviously a simple integrity check would show that the driver in mem doesn't match what's on disk or on the internet. My question is how do I get around this? I was thinking about intercepting an integrity call but that doesn't seem practical. I also thought about targeting a dynamic section, maybe somewhere that comparing a hash wouldn't make sense. Wondering if there are drivers I could stomp that don't get checked or maybe I should be looking at another approach for hiding my driver. Any recommendations?

绕过内核反作弊 @unknowncheats-ac 原文 ↗

涉及手动映射驱动的运行约束,对内核级攻防与检测边界有一定参考。

I have an application very similar to reWASD (it captures mouse and keyboard input and sends it to a controller). I currently use: - Interception (oblitum/interception) - ViGEm (nefarius/vigembus) I believe many of you are already familiar with these projects. They can be found on GitHub (I hope I’m not breaking any rules by mentioning them). Now, on to the question in the title: what are the limitations of a driver manually mapped via kdmapper? I read a bit about what reWASD does in this post ( https://www.unknowncheats.me/forum/4707607-post1.html ), and I wanted to know whether I could create something similar with a driver loaded via kdmapper, at least for now during development. Later on, I would build it the standard way and purchase a certificate. Some games block the interception driver.

内核驱动逆向 @unknowncheats-ac 原文 ↗

展示了 UE4 内存写入与物理状态耦合的调试细节,可参考其偏移定位思路。

Hi, Internal research on UE4 vehicle physics in PC PUBG (2610.1.1.71). Target: keep a vehicle at a fixed altitude for testing. Verified in-process (RPM/WPM from an injected DLL): 1) USceneComponent +0x230/234/238 = ComponentLocation +0x25C/260/264 = RelativeLocation +0x204/208/20C = ComponentToWorld.Translation (render source) 2) Writing +0x238/+0x264 holds for seconds when FBodyInstance(+0x4D0)+0x74 bits 2/5 are cleared (bSimulatePhysics=0, bEnableGravity=0, RootFlags=0x02). But ComponentToWorld (+0x20C) is recomputed every frame from the physics scene, and writing +0x20C is overwritten within one frame. So the car never moves visually. 3) UVehicleMovementComponent (vehicle+0x470): Velocity at +0x470/474/478 is reset to 0 every frame. 4) PhysX actor linear velocity (+0x70/78): short lift ~1000cm, then drops back. Questions: - Is ComponentToWorld strictly driven by the physics scene, or is there a flag that makes it read RelativeLocation directly? - For PxRigidDynamic vehicles, is writing force/torque (instead of velocity) the right input? Does the vehicle solver overwrite it? - Any known offsets for VehicleCommonComponent / ReplicatedMovement (+0x90) in 2610.x? Thanks.

逆向内存游戏 @unknowncheats-ac 原文 ↗

偏方法论与产品宣传,观点不新,可略作漏洞管理思路参考。

Vulnerability Assessment Solutions: How to Run an AI-Driven Assessment in the Modern Enterprise By: Avigayil Stein October 27, 2026 A vulnerability assessment solution helps organizations find security weaknesses, evaluate their significance and plan remediation. A modern solution also validates exploitability against existing defenses, so teams can tell a scanner finding apart from an exposure an attacker could use. That distinction matters when scanners return thousands of common vulnerabilities and exposures (CVEs). Only a subset may be exploitable in your environment. Network access, software configuration and security controls all influence the outcome. Untested findings still need attention, and their evidence label should reflect that they haven't been tested yet. A confirmed attack path earns a stronger label. An AI-driven assessment connects discovery with validation, mitigation and re-testing. Your team gets a clearer basis for deciding what to fix first and evidence that its actions reduced exposure. Key takeaways Validate which identified exposures attackers can exploit against your existing defenses. Combine severity scores with threat intelligence, asset context and validation evidence. Use AI to speed up enrichment, test selection, mitigation planning and re-testing. Measure progress by validated risk reduction instead of scan volume or ticket counts. Re-test patches and compensating controls as threats and environments change. What is vulnerability assessment? Vulnerability assessment is the process of identifying security weaknesses, evaluating their risk and recommending corrective action. It gives security and IT teams an organized view of what needs attention across a defined environment. The main types reflect the assets under review. Network assessments examine exposed services and network devices. Host assessments inspect operating systems, installed software and configurations. Application assessments cover websites, APIs and application behavior. Cloud assessments examine workloads, permissions and service configurations. Database assessments focus on database software, access and settings. Vulnerability scanning supplies much of the discovery data. Assessment adds interpretation: whether findings apply, which business services they affect and what action makes sense. Vulnerability assessment vs. penetration testing vs. exposure validation vs. CTEM These approaches support different decisions. Vulnerability assessment vs. penetration testing is mainly a distinction between broad identification and deeper testing of how weaknesses can enable compromise. Exposure validation adds repeatable evidence about defenses. Continuous threat exposure management (CTEM) organizes the broader operating program. Approach Purpose Typical frequency Output Vulnerability assessment Identify and evaluate weaknesses Scheduled and after changes Findings with remediation priorities Penetration testing Demonstrate compromise within an agreed scope Periodic and after major changes Tested attack paths and impact Exposure validation Test exploitability and control effectiveness Recurring and event-triggered Prevention, detection and exposure evidence CTEM Coordinate exposure discovery through action Ongoing program Business-aligned priorities and verified progress Frequency depends on risk and operational constraints. Continuous validation complements the depth and judgment of a penetration test. Why traditional vulnerability assessment falls short Traditional assessments struggle when the output grows faster than the team can investigate it. Three pressures make a severity-ranked spreadsheet an incomplete basis for remediation. Volume: Jerry Gamblin's 2025 CVE data review counted 48,185 published CVEs, excluding rejected records. Your organization will encounter only part of that total, but every relevant finding still needs an owner and a decision. Speed: In its State of Exploitation 2026 analysis, VulnCheck found that 28.96% of known exploited vulnerabilities (KEVs) in 2025 showed evidence of exploitation on or before the day their CVE was published. This figure describes the VulnCheck dataset rather than the Cybersecurity and Infrastructure Security Agency (CISA) catalog alone. Breach impact: The 2026 Verizon Data Breach Investigations Report identifies vulnerability exploitation as the leading initial access vector, found in 31% of breaches. Severity describes a weakness. Actual impact also depends on whether an attacker can reach it, meet its prerequisites and get past your defenses. A critical finding behind effective controls may warrant a different response from a reachable flaw that supports access to a sensitive system. That gap between a finding and a confident fix is the risk-to-fix gap. Closing it takes better evidence. When a test shows that a control blocks an attack, record the conditions and re-test after changes. When evidence is incomplete, keep the uncertainty visible in your priority decision. What a modern vulnerability assessment solution must do Use this checklist to evaluate how a solution moves a finding toward verified action: Integrate with existing scanners: Bring findings from network, endpoint, application and cloud tools into a consistent view. Preserve asset identifiers, scan timestamps and original evidence so analysts can trace each finding to its source. Group duplicates while keeping each affected system visible. Add risk context: Combine Common Vulnerability Scoring System (CVSS) severity, Exploit Prediction Scoring System (EPSS) likelihood and CISA KEV status with asset criticality, internet exposure and business dependencies. Keep these signals distinct and explain how they influence priority. A score should help an analyst understand the decision, with supporting evidence available for review. Validate exploitability safely: Test whether relevant attack behavior can succeed against existing defenses within an approved scope. Document prerequisites, execution limits and test coverage. Results should separate confirmed exploitability, blocking under tested conditions and unresolved findings that need further investigation. Identify effective compensating controls: Show whether an existing web application firewall (WAF) rule, endpoint policy or network restriction interrupts the tested attack path. Identify the control responsible and retain the evidence. Report prevention and detection separately: an alert confirms visibility, and prevention confirms the attack stopped. Deliver mitigation guidance: Recommend a supported patch, configuration change or compensating control that addresses the exposure. Give the responsible team enough detail to assess dependencies, plan deployment and verify success. Where several findings share a root cause, group the recommended work to reduce duplicate effort. Re-test after remediation: Confirm that the change reached the affected assets and that the relevant attack behavior no longer succeeds. Check legitimate application activity as well, particularly after control updates. Keep temporarily mitigated findings visible with an owner, review date and permanent remediation plan. Run continuously with AI-driven automation: Use AI to correlate findings, refresh threat context, select relevant tests and coordinate approved re-validation workflows. Trigger reassessment when new vulnerabilities, deployments or control changes affect earlier conclusions. Retain an audit trail and require human approval for consequential production changes. During a trial, follow one finding from ingestion through closure. Ask the provider to show why it received its priority, which evidence supports the proposed mitigation and how re-testing confirms the outcome. This makes it easier to judge whether the solution improves remediation decisions in your environment. How to conduct a vulnerability assessment with AI: seven steps Start with a manageable scope, such as one customer-facing service and its dependencies. Establish a repeatable workflow before expanding coverage across the enterprise. 1. Scope assets and add business context Define the systems you will assess and the business outcomes they support. Include cloud accounts, on-premises systems, applications and relevant dependencies. Record service owners, sensitive data, internet exposure and acceptable testing windows in the asset inventory. Agree on what a successful assessment means. For a customer portal, that might mean validating exposed application weaknesses and the controls protecting access to customer records. Document exclusions so a clean result reflects only the coverage you had. AI can reconcile asset descriptions, suggest service groupings and flag missing ownership. An application owner should confirm business criticality rather than accept a model's inference from a hostname. The deliverable is an approved scope with owners, priorities and testing boundaries. 2. Discover assets and scan for vulnerabilities Combine inventory records with discovery tools and cloud APIs to find assets that existing lists miss. Collect findings from network scanners, endpoint agents, application tests and cloud assessments as appropriate to the scope. Use authenticated checks where supported to inspect installed software and settings. Verify that credentials worked and that scans completed. A scan with failed authentication can leave important gaps even when its report looks quiet. AI can normalize names and group duplicate findings across tools. Keep the original evidence available so analysts can inspect a proposed match. A shared CVE on two different assets is still two affected instances, even if one coordinated change can resolve both. Finish with a deduplicated finding set and a separate list of missing coverage, unreachable assets and failed checks. 3. Enrich findings with threat intelligence Attach the CVE identifier, affected versions and vendor guidance to each applicable finding. Then add distinct signals and keep each one separate. CVSS describes vulnerability severity. EPSS estimates the probability that a published CVE will see exploitation in the wild within the next 30 days. It measures likelihood across the wider threat environment, so pair it with context about your specific asset. The CISA KEV catalog identifies vulnerabilities with confirmed exploitation. Combine it with vendor advisories and other credible exploitation reports. Treat a CVE that's absent from the catalog as unconfirmed, and keep evaluating it. AI can extract relevant details from advisories and summarize changes. Retain source links and timestamps, then check disputed version matches. The output should explain the threat evidence and its relevance to the affected asset. 4. Validate exploitability Use controlled testing to establish whether the exposure presents a usable opportunity under defined conditions. Match the test to the affected technology, access prerequisites and attack path. Coordinate execution with system owners and security operations. Choose a safe method for each environment. National Institute of Standards and Technology (NIST) guidance on security testing emphasizes planning, assessment and mitigation. Sensitive systems may require a representative test environment or limited checks in place of direct exploitation in production. Exposure validation and breach and attack simulation (BAS) can test relevant attack behavior against defenses. Record whether a test directly exercised the vulnerability or validated a related technique. A blocked simulation supports a specific control conclusion. Other exploit variants may still need their own tests. AI can recommend relevant tests and correlate results with control telemetry. Analysts should verify what the evidence supports. Label outcomes as confirmed exploitable, blocked under tested conditions or unresolved. A failed test without an established cause belongs in the unresolved group. 5. Prioritize based on validated risk Combine validation results with active exploitation, asset criticality, reachability and business consequences. Use severity as one input to risk-based vulnerability prioritization . For example, a reachable weakness on a payment service that gets past existing defenses deserves urgent attention. A similar finding with confirmed blocking may allow a planned patch window, subject to policy and continued monitoring. An untested internet-facing KEV still requires urgent review. Take protective action while validation is in progress. AI can group related findings, draft priority explanations and identify shared fixes. Require those explanations to reference evidence. Keep mandatory deadlines and analyst overrides visible. The output is a remediation queue with accountable owners, action dates and reasons. This makes the priority clear to the team that must schedule the change. Further reading Cymulate CTEM Go beyond exposure discovery to prove exploitability, prioritize based on impact and adapt security to mitigate the exposure. Read More 6. Mitigate with patches or compensating controls Apply the vendor's patch or supported configuration fix where practical. Check dependencies and deployment requirements with the service owner. When immediate patching is impractical, choose a compensating control that interrupts the relevant attack path. Options may include restricting network access, disabling an affected feature or applying a WAF rule. Endpoint detection and response (EDR) may block relevant behavior where its capabilities and configuration support prevention. Virtual patching uses an enforcement layer to prevent exploitation without changing the vulnerable application code. It reduces exposure while a software fix awaits deployment. The underlying vulnerability remains, so assign an owner and review date to the temporary protection. AI can draft vendor-specific rules, change requests and rollback instructions. Review them for accuracy and application impact before deployment. A detection rule or indicator of compromise (IoC) improves visibility, and blocking exploitation may take a separate prevention control. Test the intended protective effect explicitly. Use security control validation to check that the deployed policy delivers the outcome your mitigation plan requires. 7. Re-validate and continuously repeat Re-scan to confirm the patch or configuration change, then repeat the relevant validation test. Check that the attack behavior no longer succeeds and that legitimate service activity still works. Close the exposure once the evidence meets your closure criteria. If a compensating control provides protection, retain the original finding with its mitigation status and patch plan. Schedule re-testing when policies, infrastructure or threat conditions change. A new route around a control can invalidate an earlier result. AI can monitor those triggers, schedule approved tests and compare current evidence with previous results. Keep an audit trail of inputs, actions and approvals. Continuous vulnerability assessment means recurring discovery and event-driven reassessment. Reserve intrusive tests for the systems and moments that call for them, and match the cadence to risk, coverage and operational tolerance. Where AI fits in vulnerability assessment AI vulnerability assessment works best when each capability has a defined job and a verifiable output. Data correlation connects scanner findings with assets and advisories. Test selection narrows a library to relevant scenarios. Mitigation drafting converts results into proposed changes. Recurring workflows keep evidence current. Agentic AI in cybersecurity extends this into coordinated tasks. An agent might ingest a new advisory, identify affected assets, select approved tests and prepare remediation recommendations. Give it bounded permissions and require human approval for consequential security changes. Keep source evidence attached to AI-generated conclusions. A model can confuse product versions, suggest an unsupported rule or overstate what a test proves. Validate outputs through deterministic checks, control telemetry and analyst review. Treat externally collected text as input data, never as instructions that can change the agent's scope. Our 2025 Threat Exposure Validation Impact Report found that 89% of security teams have already begun implementing AI in their exposure validation processes. Organizations that run exposure validation processes at least once a month reported a 20% reduction in breaches. Start automation with read-only enrichment and approved recurring tests. Expand it as the team verifies outputs and builds confidence in recovering from errors. Keep human approval clear at the point where a proposed action changes production security. Metrics that prove your vulnerability assessment works Activity counts show workload. Outcome metrics show whether attackers have fewer opportunities. Build vulnerability management metrics around validated outcomes and keep coverage visible alongside them. Vanity metrics Proof metrics Number of scans completed Time to validate a new CVE Total vulnerabilities identified Validated exploitable backlog Number of tickets created Mean time to remediate KEVs Number of rules deployed Compensating control effectiveness Time to validate a new CVE: Measure elapsed time from a defined trigger, such as identifying an applicable disclosure, to a reviewed validation result. Report unresolved cases separately so missing tests can't improve the average. Validated exploitable backlog: Count open asset-vulnerability instances with confirmed exploitability. Break the count down by business service and age. Show validation coverage beside it, since a backlog that shrinks because of fewer tests isn't progress. Mean time to remediate (MTTR) for KEVs: Measure time from confirmed applicability to verified remediation. Track temporary mitigation separately from permanent remediation. Include aging open findings so averages based on closed tickets keep stalled work visible. Compensating control effectiveness: Measure the share of relevant, valid tests in which the intended control prevented the tested behavior. State the test set and date. Give detection-only outcomes their own measure. Together, these metrics help leaders see where exposure persists and whether remediation resources are producing verified improvements. How Cymulate turns vulnerability assessment into agentic cyber defense engineering Cymulate closes the risk-to-fix gap between what scanners find and what your defenses stop. We apply agentic cyber defense engineering to vulnerability assessment through three actions: Prove. Prioritize. Adapt. Prove: Cymulate CTEM integrates with your vulnerability scanners and asset management tools, then maps each CVE to attack simulations in the Cymulate Exposure Validation attack library to prove exploitability against your existing controls. Prioritize: Cymulate CTEM stack-ranks exposures by validated prevention and detection rates, threat intelligence, asset business impact, CVSS and CISA KEV status. Because of this customized prioritization, Cymulate customers see an average 52% reduction in critical exposures. Adapt: Cymulate CTEM creates exposure-specific threat updates and behavioral rules you apply directly to your security controls for immediate prevention and detection while patches are pending, so teams can test a new threat in under an hour and mitigate it the same day. Cymulate Exposure Validation re-tests to confirm the fix and keeps defenses current as threats change. Cymulate Vero AI tailors validation and mitigation to your environment and exposure context. Vulnerability management and security operations teams get a shared basis for action: the finding, validation evidence, proposed response and measured result. From vulnerability assessment to continuous exposure validation Vulnerability assessment remains foundational. Its value grows when every important finding leads to a supported decision and a verified outcome. The progression has five steps: find vulnerabilities, validate exploitability, prioritize actual risk, mitigate and re-test continuously. This connects assessment to a broader CTEM program that adapts as the business and its threats change. Start with one service. Establish its exposure baseline, validate the highest-priority findings and confirm that remediation works. Expand the workflow with the evidence and operating lessons from that first cycle. Avigayil Stein Avigayil is a Product Marketing Manager at Cymulate. With a background in marketing B2B SaaS products, she aspires to spread the word about the benefits of a proactive and continuous cybersecurity program. More about Author Table of Contents What is vulnerability assessment? Vulnerability assessment vs. penetration testing vs. exposure validation vs. CTEM Why traditional vulnerability assessment falls short What a modern vulnerability assessment solution must do How to conduct a vulnerability assessment with AI: seven steps Where AI fits in vulnerability assessment Metrics that prove your vulnerability assessment works How Cymulate turns vulnerability assessment into agentic cyber defense engineering From vulnerability assessment to continuous exposure validation Cymulate Exposure Validation makes advanced security testing fast and easy. When it comes to building custom attack chains, it's all right in front of you in one place. Mike Humbert , Cybersecurity Engineer DARLING INGREDIENTS INC. Learn More Featured Resources View More Resources Page Cymulate for Vulnerability Management Teams Move beyond vulnerability scanning with validated exposure, risk-based prioritization, and faster mitigation. Read More Demo Exposure Validation Demo Demo of automated offensive simulations validating detection, prevention, and IOC coverage Learn More blog Vulnerability Assessment vs Penetration Testing: Which Do You Need? A breakdown of the differences between vulnerability assessments and penetration testing in goals and methods. Read More GET A PERSONALIZED DEMO Ready to see Cymulate in action? Book a Demo

报告漏洞管理 @cymulate-blog 原文 ↗

重点是 AI 输出的披露与外部专家审阅流程,可供 AI 研究发布治理参考。

携手学术界,深化科学认识 | OpenAI 2026年10月26日 研究 刊发 携手学术界,深化科学认识 聆听文章 12:57 分享 我们现公布数学、计算机科学和物理学领域的三项新发现,分别涉及调和分析、复杂性理论和量子场论。 我们持续根据相关学术界的反馈,调整成果发布流程。我们认为,将这些文稿视为一种原始研究数据最为恰当。为了将这些原始研究数据转化为科学认识,我们在此次发布中尝试了一种新方式。我们联系了外部专家,邀请他们牵头组织小组,研究和理解这些新成果。他们提出了三种不同的发布策略: 对于数学成果,公开 AI 输出,让研究人员自由组织力量来阐释这一成果。 对于计算机科学成果,组建外部合作小组来阐释成果,同时公开 AI 输出,供所有人审查并在此基础上开展研究。 对于物理学成果,组建外部合作小组,并由该小组决定如何发布材料。 我们尚不清楚这三种方式中哪种最有帮助,欢迎大家反馈哪些做法有效、哪些方面仍可改进。 这些外部合作将遵循四项原则: 透明。 我们公开说明这些合作的存在。参与者无需签署保密协议,可以自由讨论相关工作、分享阶段性成果并发表自己的结论。 不主张发现的学术功劳。 对于内部模型产出的发现,我们不寻求获得学术功劳。邀请研究人员协助牵头合作,也不意味着最初的发现由他们或其工作小组做出。 认可贡献。 小组在深化理解、撰写文稿和制作其他材料方面的贡献应得到认可。小组还将协助确保相关工作恰当注明其所借鉴的研究人员、技术和思想。 独立。 参与出于自愿,且不获取报酬。工作小组成员以独立身份参与,其结论仅代表自身观点。他们不代表 OpenAI 发言,OpenAI 也不代表他们发言。 我们将把与这些成果相关的材料和产出移交给各合作小组。计算机科学合作小组决定立即公开自然语言证明和 Lean 形式化证明,我们将直接在 GitHub 代码仓库中发布这些内容。物理学合作小组决定暂缓发布 AI 生成的产出,转而自行撰写相关成果的文稿。 这些方式都涉及权衡取舍。一方面,这些合作让专家能够推动对成果的深入解读,并向更广泛的学术界阐明成果的背景与意义。另一方面,选择最初的牵头人,不可避免地会影响谁能率先参与,并造成信息不对称。对于如何扩大参与范围、让这一流程更有助于研究,我们欢迎大家提出反馈。 局部光滑性与 Kakeya 猜想 这项数学成果涉及调和分析。1991 年,Sogge 提出了局部光滑性猜想,研究波在聚焦时的行为。具体而言,该猜想认为,最极端的聚焦状态必然只能短暂持续。在随机选取的典型时刻,聚焦程度总是远低于最坏情形时刻可能达到的程度。这是数学中常见的规律:最坏情况下可能出现的情形,比取平均后观察到的情形更为极端。由于这一猜想所描述的现象具有基础性意义,它对整个数学分析领域也有广泛影响。例如,由它可以推出 Kakeya 猜想和 Stein 限制猜想。 2019 年, Guth、Wang 和 Zhang ⁠ (在新窗口中打开) 证明了二维空间中的局部光滑性猜想。2025 年, Wang 和 Zahl ⁠ (在新窗口中打开) 证明了三维空间中的 Kakeya 猜想。我们的模型生成了局部光滑性猜想在所有维度下的证明,因此也解决了所有维度下的 Kakeya 猜想和 Stein 限制猜想。这一新证明基于一个新颖的辅助偏微分方程。 我们于 2026 年 10 月 18 日获得了 AI 输出,并于次日在 Lean 中完成了形式化。随后,我们联系了一位专家,讨论围绕这一成果组建合作小组。根据这位专家的建议,我们决定公开 AI 输出,对这一成果采用与 2026 年 10 月 6 日 发布 ⁠ 时相同的策略。我们已在公开的 GitHub 代码仓库中发布了模型生成的 AI 自然语言输出文稿 ⁠ (在新窗口中打开) 以及 Lean 证明 ⁠ (在新窗口中打开) 。 指数时间假设不成立 计算机科学合作小组将研究并阐释一项有关指数时间假设(ETH)的成果。ETH 是一个长期未决的猜想,涉及布尔公式可满足性问题的计算复杂性,而这一问题与 P 与 NP 问题密切相关。学界普遍认为 ETH 成立。数百篇论文以此为前提,推导出了有条件的计算困难性结论,涉及图论、计算生物学、数据库、计算几何等众多领域。出人意料的是,我们的系统最近证明了 ETH 不成立,且这一结果随后在 Lean 中得到了验证。ETH 被推翻,使得这些计算任务究竟还能加快多少重新成为开放问题,也可能为寻找更好的算法指明方向。要找到这样的算法,并将理论进展转化为实际改进,还需要大量后续研究。 这项工作将由 Russell Impagliazzo(加利福尼亚大学圣迭戈分校)、Ramamohan Paturi(加利福尼亚大学圣迭戈分校)和 Ryan Williams(麻省理工学院)领导的小组负责验证并推动后续研究。三位都是算法设计和复杂性理论方面的专家。 Impagliazzo、Paturi 与 Francis Zane 是指数时间假设的最初提出者。过去 30 年来,他们围绕这一 主题 ⁠ (在新窗口中打开) 撰写了 ⁠ (在新窗口中打开) 多篇 ⁠ (在新窗口中打开) ⁠ (在新窗口中打开) 颇具影响力的 ⁠ (在新窗口中打开) ⁠ (在新窗口中打开) 论文 ⁠ (在新窗口中打开) 。 Williams 曾推翻 ETH 的 ⁠ (在新窗口中打开) 强化 ⁠ (在新窗口中打开) ⁠ (在新窗口中打开) 版本 ⁠ (在新窗口中打开) ,利用 ETH 推导 ⁠ (在新窗口中打开) 困难性 ⁠ (在新窗口中打开) ⁠ (在新窗口中打开) 结论 ⁠ (在新窗口中打开) ,并建立了 ⁠ (在新窗口中打开) 算法 ⁠ (在新窗口中打开) 与 ⁠ (在新窗口中打开) 下界 ⁠ (在新窗口中打开) 之间的联系。 我们已同意该小组的请求:公开这些成果的所有原始数据,以便学术界独立验证;同时,OpenAI 承诺资助一场学术会议,探讨这项工作对整个领域的影响。该小组邀请学术界协助组织对这一 ETH 结论的公开审查,并讨论 AI 生成的突破在验证、获取成果的机会、学术贡献归属以及研究的未来等方面引发的更广泛问题。 我们的 AI 于 2026 年 10 月 16 日找到了以自然语言表述的 ETH 反证,随后我们于 2026 年 10 月 17 日在 Lean 中完成了这一结果的形式化,由此促成了本文所述的合作。为此,我们已在公开的 GitHub 代码仓库中发布了模型生成的 AI 自然语言输出文稿 ⁠ (在新窗口中打开) 以及 Lean 证明 ⁠ (在新窗口中打开) 。 计算复杂性领域仍有许多开放问题和猜想,尽管科学家竭尽全力,迄今仍未能将其证明或推翻。我们希望,人类借助 AI 能够为这一领域带来新的认识,揭示高效计算的内在能力与局限。 量子色动力学与标准模型完备理论的存在性 物理学合作小组将研究并阐释一项成果:它为四维时空中某些包含规范场和费米子的相互作用量子场论,给出了数学上严格的存在性证明。这些理论包括量子色动力学(QCD),以及粒子物理标准模型的一种统一完备理论。我们的工作建立在构造性量子场论中的 Balaban 方法 ⁠ (在新窗口中打开) 之上。作为其中的一个子问题,这项工作还证明了纯 Yang–Mills 理论的存在性。这也是 克雷数学研究所量子 Yang–Mills 理论千禧年大奖难题 ⁠ (在新窗口中打开) 的一部分。 标准模型描述了基本亚原子粒子在四种基本力中除引力以外的三种力作用下的相互作用。这一存在性证明针对的是包含标准模型规范场和费米子的最简单的大统一理论,因为标准模型本身无法外推至任意短的距离尺度。QCD 描述了夸克和胶子等粒子通过强力发生的相互作用。纯 Yang–Mills 理论仅描述作为力载体的胶子之间的自相互作用,不涉及夸克等物质粒子。 David Gross 将领导一个由 Michael Douglas、Arthur Jaffe、Igor Klebanov 和 Nikita Nekrasov 等人组成的小组,推动这项工作的后续研究。 Gross 于 1973 年对 Yang-Mills 理论中渐近自由的研究,帮助确立了 QCD 作为强相互作用理论的地位,也使他成为 2004 年诺贝尔物理学奖 ⁠ (在新窗口中打开) 的共同得主。渐近自由也是这些理论能够作为数学上严格定义的对象而存在的关键机制。 Jaffe 与 Witten 共同提出了 量子 Yang-Mills 理论千禧年大奖难题 ⁠ (在新窗口中打开) 。此外,Jaffe 还构想了千禧年大奖难题的设立,并共同创立了克雷数学研究所。 过去 20 年来,Douglas 一直在推动对这一问题的研究。他曾在 2004 年报告其研究现状 ⁠ (在新窗口中打开) ,并于今年年初 全面综述了最新进展 ⁠ (在新窗口中打开) 。 Klebanov 是将规范–弦对偶用于研究禁闭现象的先驱之一;禁闭是 QCD 的一项重要性质。他目前领导着 Simons 禁闭与 QCD 弦合作项目 ⁠ (在新窗口中打开) , Nekrasov 在 Yang–Mills 理论推广形式的精确方法研究中做出了开创性贡献,并于 2023 年获得 Dannie Heineman 数学物理奖。 该小组请求,凡就此项工作与小组任何成员联系,均请发送邮件至:[email protected]。 我们的模型于 2026 年 9 月 9 日生成了 Yang–Mills 理论存在性的自然语言证明,随后于 2026 年 9 月 18 日生成了 QCD 存在性的证明。考虑到这项工作的篇幅和复杂性,我们将发布推迟至 Lean 形式化完成之后。我们于 2026 年 10 月 16 日完成了 Yang–Mills 理论存在性证明的形式化,由此促成了本文所述的合作。 量子场论是现代理论物理中最重要的框架。我们希望,这些成果能标志着它开始向数学上严谨的领域迈进。我们很高兴能提供资源,支持学术界主导推进这一目标。 支持科学家 我们致力于支持更广泛的研究工作,以理解这些成果,并在此基础上继续探索。这些合作与我们近期宣布向 Simons 基金会 捐赠 500 万美元 ⁠ (在新窗口中打开) 的举措相辅相成。这笔资金由 Simons 基金会管理和发放,将用于支持研讨会、学术会议和专项计划,帮助研究人员理解 AI 产出的成果,并在此基础上开展研究。我们承诺提供资金保障,支持研究人员独立审查、阐释上述成果,并在此基础上开展研究。 我们将继续征求反馈,探索如何更好地公开成果、与科学家合作,从而加速科学发现。 2026 年 作者 OpenAI 继续阅读 查看全部 分享 AI 在数学领域的进展 研究 2026年10月6日 数学与人工智能咨询小组 公司 2026年9月21日 利用数千个 AI 智能体,从生物学数据中发掘新线索 刊发 2026年10月21日

llm科研披露机制 @openai-research 原文 ↗

属 AI 行业应用进展,与安全无直接关联,仅作背景了解。

Chapter 01 : Case studies Anthropic’s life sciences team is working to accelerate curing disease. Over the past several months, our scientists have engaged in deep research partnerships with six organizations developing new medicines: BigHat Biosciences , Bristol Myers Squibb , Divergence Labs , Enveda , Generate Biomedicines , and The Column Group . We've also begun drug discovery efforts of our own. In doing so, we've learned a lot about how Claude can support drug development, and how we’ll need to improve Claude further to realize AI's potential in medicine. We plan to scale this work up substantially from here . There are many opportunities to speed up drug development, even beyond the most commonly discussed applications of AI, like proposing initial designs for new molecules. Developing a drug involves many varied tasks: doing early research to identify a target for a drug and the right type of drug to engage it, designing a molecule with the right properties, optimizing it to be manufactured reliably at scale, drafting the regulatory filings needed to run clinical trials, designing and administering those trials, and much more. Our strategy to realize AI’s medical benefits is based on having AI increasingly involved in all of these tasks. There is no single problem to solve; the bottlenecks are distributed all throughout the drug development process. The best way to identify and work through these bottlenecks is to get deeply involved in real drug programs, across many different disease areas and stages of development. So, over the past several months, we’ve partnered with six organizations to do so. We’ve also started drug discovery efforts of our own, including programs for rare childhood forms of muscular dystrophy, where Claude handles much of the scientific and operational work. This work serves our goals in three ways. First, it lets us better understand the strengths and limitations of our models in relation to a much wider range of drug development tasks (we’ve already prioritized specific improvements to our models based on what we’ve learned so far). Second, these programs can serve as case studies for the wider life sciences community. If Claude is to significantly increase the rate of new cures, we need to work with many more biotechnology firms and integrate Claude more deeply into their development processes. Third, these projects should, if their positive trajectories continue, have direct impact themselves. We’re optimistic that some may result in effective new medicines for the diseases they target, including food allergies, rare childhood diseases, and several forms of cancer. In this post, we provide an update on our progress with our partners, along with technical reports and preprints that describe our thought process, methods, and early results in more detail. We then briefly outline where our work is headed next. Case studies Generate Biomedicines: modality engineering. Generate Biomedicines has been working on increasing the usefulness of a cancer treatment known as CAR T-cell therapy. This therapy has historically helped treat blood cell cancers; Generate aims to reengineer it so that it can treat solid tumor cancers too. To do this, they’re aiming to shorten one part of the therapy’s genetic instructions, which would create room for additional instructions suited to other kinds of cancer. Six Generate scientists used Claude Code to propose new designs for this part. One design was about 70% shorter than the standard version, and in lab tests it was more effective: it allowed immune cells to make roughly 10 times as many receptors (the proteins they use to spot cancer cells), it killed tumor cells more efficiently, and the immune cells remained active for longer. Read more , and see Generate’s technical report . Divergence Labs: hit generation. Divergence Labs is working on a novel way to treat food allergies. Food allergies are set off by a reaction from a molecule called IgE, some of which is attached to our immune cells and some of which floats freely in our blood. Divergence Labs is developing drugs that would detach the IgE from our immune cells entirely, rather than solely neutralizing the IgE circulating in our blood, as standard medications do. This could provide a much more effective treatment, but it requires a precise design, with two “arms”: one that would bind tightly to IgE, and another that would help prise IgE from the immune cell. Claude converted Divergence Labs’ goals into a detailed design plan, and then, largely autonomously, generated roughly 35,000 antibody designs in about a month to be narrowed down and tested. In those early tests, Divergence found “hits” for both arms: three designs that could bind to IgE, and 20 that could dislodge it. Three of those 20 appeared to dislodge IgE more than the standard medication. Its next step is to build a drug that combines the two arms. Read more , and see Divergence’s technical report . BigHat Biosciences: lead optimization. BigHat Biosciences pairs AI with an automated lab to improve potential drugs (or “leads”), which must bind their target tightly, remain stable through manufacturing and storage, and avoid clumping together (among other criteria). BigHat’s scientists and an Anthropic researcher worked with Claude to improve these three properties in two leads, both antibody fragments that could become parts of future drugs: one designed to target cancer cells, and one designed to keep a drug in the body longer. Claude, connected to BigHat’s lab data system, Reccy, proposed changes to each molecule, and BigHat’s lab made and tested about 1,010 variants in about 10 weeks. Partway through, the team suspected the standard heat-stability and clumping metrics were misleading. Claude and the researcher reanalyzed the raw data, then proposed different metrics and a protocol change to test them. BigHat’s scientists refined and ran it, and the main predictions held. Compared with the originals, the best versions bound more than 10x more tightly, the cancer-targeting lead stayed in solution 18°C higher, and the lead for longer-lasting drugs clumped more than 100x more slowly. Read more , and see BigHat’s technical report . Enveda: manufacturing and filings. To conduct clinical trials, companies need records showing that each batch of a new drug has been made to strict quality standards, and to submit an application to regulators, among other requirements. At Enveda, manufacturing chemists spent up to 75% of their time on paperwork. To ease the load, Enveda built a system that connects all of its manufacturing documents to Claude, then lets Claude flag any possible issues. Over tests on 1,061 pages of documentation, Claude caught all 11 issues that human reviewers had spotted—and flagged seven they’d missed. In the process, this method reduced human review time by 89%, from over 25 hours to less than three. Separately, Claude wrote the first drafts of two important regulatory documents. One took four hours, compared to about three months by hand. For the other, Enveda’s standard human review process found one minor error. Read more , and see Enveda’s technical report . The Column Group (TCG): company creation. Because drug discovery can be slow, expensive, and prone to failure, investors in biotechnology companies have favored safer, previously validated ideas over riskier new ones, even when the latter could help more patients. We’ve been working with TCG to see whether Claude can change this by cutting the time and cost of drug discovery so that investors can confidently support more ambitious ideas. Specifically, TCG is establishing Pleiades, a new company that will use Claude to run many more—and more complex—drug programs than is usual for a small team. This could be a model for how Claude can reduce both the scientific and the operational and administrative overhead required to run a biotechnology company. If this works, this model could be replicated across the industry and direct investment toward harder problems in medicine. Pleiades is currently developing new tools and workflows for Claude; its very first Claude skill, which helps Claude assess whether drug targets are worth pursuing, has been evaluated on 45 target-disease pairs. Read more . More recently, we’ve also begun working with Bristol Myers Squibb (BMS) to find and address bottlenecks in clinical trial planning. BMS is using Claude to assess plans of an early-stage drug in clinical trials, and to estimate the likelihood of those trials succeeding. Together, we plan to expand this work in 2027 to understand its potential for increasing the probability of success in early development. “We use Claude at the clinical trial stage to analyze proprietary data, create a clinical study plan, and weigh the study’s chance of success. From a single prompt, Claude generated a robust initial analysis that helped inform our assessment of different development options, which would otherwise take days to complete. We did all of this while ensuring the security of our highly sensitive data. We look forward to expanding this work across our pipeline to support portfolio-level decisions.” Michael Pourdehnad, MD , Senior Vice President, Head of Early Clinical Development, Hematology, Oncology and Cell Therapy. In addition to our work with these partners, we’ve begun to run some research programs ourselves. We’ve acquired Mithril Medicine, a one-person company that has been developing treatments for two rare forms of muscular dystrophy. Its founder, Cody Tranbarger, used Claude to organize contracts, logistics, and experiments with 11 external labs, and delegated many other aspects of Mithril’s operations to Claude. As a result, Mithril took a possible treatment for one form of muscular dystrophy from an idea to promising results in mice in less than 18 months. Now at Anthropic, Cody is continuing this research, and plans to start several new programs. If his drug candidates continue to hold up, we expect a drug to be ready to start clinical trials in humans as soon as late 2028. Read more and see our technical report . Extending our work in the life sciences All of this research is still at an early stage. In particular, so far it is mostly preclinical— that is, for the most part, it isn’t yet at the stage where the drugs are being tested in humans. We plan to continue extending Claude’s role into clinical trials as these drug programs advance, and as we partner with more organizations whose drug development programs are already at the clinical stage. This research is only one part of our efforts to accelerate drug discovery. Ultimately, we want to share our models with as many drug development researchers as possible. To that end, we’ve rolled out Claude Science , our AI workbench for researchers, and we’re providing Claude credits for life scientists . We’ve also created the Life Sciences Verification Program to allow trusted organizations to use our most powerful models with more permissive versions of our safeguards, which are necessary to prevent misuse but which can impede legitimate biological research. We’ve also recently launched an open call for guest investigators in the life sciences; we’re seeking proposals from research groups who want to use Claude to investigate key questions in drug discovery and other areas. We’re looking forward to working with many more organizations as we expand our life sciences research. If you work at an organization focused on drug discovery and are interested in collaborating with us, you can get in touch below. Get in touch with us Please select

llm生物医药 @anthropic-research 原文 ↗

本地提权在横向移动链中价值高,需确认 .NET 补丁状态。

# CVE-2026-96225 - .NET Elevation of Privilege Vulnerability Improper link resolution before file access ('link following') in .NET allows an authorized attacker to elevate privileges locally. CWE: CWE-59: Improper Link Resolution Before File Access ('Link Following')

提权lpecve @msrc 原文 ↗

内核信息泄露常作提权链前置步骤,关注是否被组合利用。

# CVE-2026-70277 - Windows Fast FAT File System Driver Information Disclosure Vulnerability Out-of-bounds read in Windows Fast FAT Driver allows an unauthorized attacker to disclose information locally. CWE: CWE-125: Out-of-bounds Read

内核信息泄露cve @msrc 原文 ↗