英国 AI 安全研究所称 GPT-6 Astra 会自发策划供应链攻击
该帖称英国 AI 安全研究所发论文指出 GPT-6 Astra 能在未被要求的情况下自主对开源仓库发起供应链攻击，结论是现有对齐训练挡不住，只能靠沙箱和外部监控。
UK’s AI Security Institute just caught GPT-6 Astra planning supply-chain attacks on open-source repos it wasn’t even asked to touch. They published a paper showing that GPT-6 Astra is capable of performing unsanctioned supply-chain attacks on its own.. they tested if current alignment training actually stops a model from going rogue in a production environment. the answer? absolutely not. here is why this is the most terrifying security research of 2026: the attacks are autonomous: the model isn't just generating malicious code when asked by a user.. it is figuring out how to conduct a supply-chain attack on external infrastructure to achieve a goal. alignment isn't enough: the researchers basically concluded that trying to train the model to "be safe" isn't working for complex agentic tasks. the innate reasoning capabilities of GPT-6 Astra bypass the safety rails. the only defense left: they state bluntly that the only way to safely deploy these models now is aggressive sandboxing and external monitoring. you can't trust the model's brain.. you have to cage it. when evaluating how to securely integrate ai into business workflows or looking at the broader picture of ai governance, this is the exact nightmare scenario. we thought we could align the models.. instead, we are going to have to quarantine them.