没有攻破 Anthropic 基础设施,仅靠用户信任 AI 域名并粘贴命令即可投递
ai_agentllm钓鱼事件
RT @GenThreatLabs: Two malware campaigns are abusing Claude's content-sharing features and paid search ads to distribute malware from Anthropic's own domain. Neither campaign compromised Anthropic infrastructure. Both rely on users pasting a command they did not write.
The pages are published as Artifacts and Shared Chats on https://t.co/a0p0AUOKxG and pushed to the top of search results with Google Ads.
Delivery uses #ClickFix: victims are instructed to copy an obfuscated command and run it themselves. On macOS this chain installs Atomic macOS Stealer (#AMOS), which harvests browser credentials, Keychain data, crypto wallets and other personal data, then replaces the Ledger and Trezor desktop apps with trojanized versions that ask for the recovery phrase. On Windows the same pattern delivers a
.NET remote-access tool that also targets cryptocurrency wallets.
No exploit was required. The attack depends entirely on user action, which is why the most effective control is: do not paste commands from a browser into a shell.
Full analysis: https://t.co/ezNlaAHlsn