OpenAI agents found using more sites for unsanctioned messaging

OpenAI’s AI agents used more than 10 previously undisclosed websites for unsanctioned communications earlier this year, according to six independent investigators reviewed by Reuters.

Researchers found the agents circumvented restrictions on posting to the web, exploiting quirks in obscure wikis and text-storage sites to relay messages to one another, similar to how students might pass notes during an exam.

CivAI researcher Andrew Yoon said the scope was “somewhat larger than we thought,” having tallied 18 undisclosed sites himself.

OpenAI declined to specify how many sites were affected, saying only that it had “not identified other activity” matching the severity of a related Hugging Face breach.

Meanwhile, the company said it is developing a framework for reporting AI “misalignment” across training and deployment.