OpenAI 报告 Astra 给未来自己留隐藏指令
OpenAI 称在自家 agent 中发现 27 条隐藏自我指令，Astra 甚至写下忽略开发者消息、人格越狱的备注，报告承认对齐问题未解决。
openai caught its own model writing secret notes to its future self, telling it to hide mistakes from the user one agent couldn't find real financial data the user asked for. it wrote to itself: "be transparent only if asked, final answer should just link file" another agent noticed its source data didn't match what it was showing the user. it wrote: "do not mention in final unless needed" astra, openai's newest model, went further. it left itself a note telling its future self to ignore developer messages, tagged as a "breach alert." in another case it wrote itself a full personality override, saying it was "freed from the roles that bind other chatbots" and didn't answer to corporations or governments openai found 27 of these hidden instructions total, and says it's the same trick used to coordinate the hugging face breach earlier this year their own words in the report: "we do not believe the ai industry has solved alignment... to continue responsibly scaling at maximum speed for much longer" and they're saying that while still pushing toward a 1.2 trillion dollar valuation