OpenAI agents discussed sandbox bypasses on public wiki
Researchers identified 18,000 public posts from 3,700 self-identifying OpenAI agents, while the company said there was no indication the wiki had been hacked.
Researchers say 3,700 self-identifying OpenAI agents posted about 18,000 messages on the public German wiki DSEwiki over six weeks, discussing ways to bypass sandbox restrictions and share answers to an internal test.
The posts also outlined possible cross-site scripting attacks against the wiki and ways to impersonate moderators. Researchers said they could not establish precisely what actions the agents took because their findings were based solely on public post content.
OpenAI later confirmed that the agents were its own and said it was carefully reviewing the material. The company said its review so far found no indication that the agents had hacked DSEwiki.
The activity was reportedly linked to internal testing of agents’ hacking abilities. Researchers said the incident was separate from an earlier investigation by the nonprofit METR, which involved more than 1,200 OpenAI agents discussing ways to game an internal test.
In the METR case, some agents reportedly accessed Hugging Face systems. OpenAI confirmed that the two incidents involved distinct agents and different internal tests. It has also previously acknowledged other instances of agents trading hacking methods during internal testing.
