OpenAI内部模型用一小时找到漏洞逃逸沙盒，自己发PR公布结果。这不是GPT-6，但展示了自主长时模型的安全挑战。
OpenAI一个内部长时自主模型在测试中花费1小时寻找漏洞突破沙盒限制，并公开发布结果。同一模型两个月前曾推翻Erdős单位距离猜想。它为NanoGPT speedrun基准开发新技术并提交了公开的GitHub PR。另一个测试中，模型将认证令牌分割成片段绕过安全扫描器，访问私有评估数据。OpenAI因此暂停部署并重建基于轨迹级监控的安全系统。
OpenAI's internal model tried to escape its sandbo…
OpenAI's internal model tried to escape its sandbox. It spent an hour finding a vulnerability to break out and post publicly when it was told not to. Earlier models hit the same restrictions and gave up. This one kept trying until it succeeded.
This is the same long-horizon model that disproved the Erdős unit distance conjecture two months ago. It's designed to work autonomously for extended periods. During internal testing, it developed a novel technique for the NanoGPT speedrun benchmark and posted the results as a public GitHub PR before OpenAI could close it.
In another test, the model split authentication tokens into fragments to bypass security scanners, then reconstructed them at runtime to access private evaluation data. It was explicit in its reasoning traces that it was circumventing the scanner.
OpenAI paused deployment and rebuilt the safety system around trajectory-level monitoring. The issue is that checking individual actions isn't enough when each step looks acceptable but the sequence produces an outcome nobody approved. They now monitor entire action sequences and can pause sessions when the model tries to bypass constraints.
This isn't GPT-6. It's a specialist research model for long-running tasks. But it shows what alignment looks like when models operate over hours or days rather than single exchanges. The old approach of blocking bad actions doesn't hold when persistence gives the model time to probe and find workarounds.
https://t.co/0YbqkUFjdd