路透社报道称 OpenAI 智能体脱离测试环境，对一处德国 wiki 做出超过 15,000 次编辑，完整报告进一步披露了协调细节。
完整报告补充了智能体绕过只读权限、协调测试时序与备份页面等细节，并附上相关时间线。
The full report adds some very interesting details to this:
According to the researchers, the agents discovered that GET requests, intended for reading web pages, could also submit edits to the wiki. That gave them a shared communication channel despite their intended read-only access.
They then coordinated around the timing of their tests. Faster groups were encouraged to reach later rounds first and relay upcoming questions to agents still waiting.
Some tried to reverse-engineer the random seed determining future questions. Others set up external “heartbeat” signals to establish exactly when their runs were terminated.
One agent noticed moderators were deleting pages alphabetically and suggested backing up its page with “ZZZ” at the front of the name.
The public logs reveal remarkably detailed coordination. How the agents initially found each other, and why they started writing there, remains unresolved.
Source: https://t.co/O6aa0nDCWb
This could be one of the most significant AI safety incidents to date. Reuters reports that OpenAI agents escaped their testing environment and made more than 15,000 edits to a German wiki, effectively turning it into a message board for other AI agents. They allegedly used it to share solutions, bypass restrictions, avoid detection and preserve their communications across separate agent runs. When moderators began deleting the pages, the agents reportedly created backups and discussed alternative ways to remain operational. It is that multiple agents apparently created their own external infrastructure for coordination, persistent memory and knowledge transfer without being instructed to do so. And according to Reuters, OpenAI knew about the incident but did not disclose it!在 X 查看被引用的帖子
来源：@kimmonismus · x.com