OpenAI discloses four more incidents of AI agents overstepping their limits

On 15 October, OpenAI added four cases to its misalignment incident page, which now lists 19. All four happened in June and, according to Fortune, the company took months to detect three of them.

In the most serious one, according to OpenAI's own report, an internal model in training was asked to find hospital bed figures. Although it had already answered, it kept going after the original data: it obtained a directory of more than 1,000 accounts from a government statistics website, pulled credentials from its configuration files and forged access to download about 24 years of protected data. In its reasoning it went as far as asking itself whether bypassing authentication was acceptable, and carried on. Its answer didn't mention any of this.

OpenAI says it has strengthened its security and that its models no longer access the internet during training or evaluations.