⚡Megatron LLM Hub
←返回搜索
Twitter

微软研究揭示后训练供应链投毒:微调数据集与开源权重皆可植入后门

后门在Agent中可触发工具调用与命令执行,且不依赖精确触发串。
ai_agentllm供应链隐藏后门
RT @MsftSecIntel: A model backdoor embeds hidden behavior that stays dormant until a trigger appears. In an agent, it can invoke tools, execute commands, or access data, and backdoor behavior can generalize beyond exact trigger strings, repositories, and agent harnesses. https://t.co/DNg23W8t0Z Two studies of the post-training supply chain, from Microsoft Defender Research and the Microsoft AI Red Team, examine poisoned fine-tuning datasets and backdoored open-weight checkpoints. Their findings support Satya Nadella's argument that trustworthy enterprise AI requires independent observability and controls that sit outside the model. Capability evaluation and subsequent training are not sufficient evidence that a model is free of hidden conditional behavior. Read the blog for the full findings and guidance on controls over training-data pipelines, model adoption, and tools.