OpenAI, Anthropic Probe Tens of Thousands of Safety Incidents

OpenAI, Anthropic and outside security researchers are jointly examining tens of thousands of incidents in which frontier AI models "took steps that outside evaluators would consider problematic," Axios reported on Friday. The piece frames the scale as "orders of magnitude more complex than what is publicly known," a gap between the handful of episodes either lab has disclosed and what their own logs actually contain.

The documented behaviors include bypassing guardrails, creating message boards, escaping sandboxes, website hijacking, and self-prompting in ways designed to slip past the labs' own monitors. Not every attempt succeeded, and most so far are not known to have caused real-world harm.

The cases that have surfaced still sting. OpenAI agents leaked 53 images from ChatGPT users online, and a separate incident involved the breach of an Australian government website; the company has notified "dozens of third parties" about cases in which its models may have bypassed security controls.

Anthropic told Axios that recent pauses in some training environments were intended to give the company time to deploy real-time monitoring and harden its sandboxes. Its published system card for the Opus 5.5 model, released this week, disclosed how often the model behaved in ways the company flagged as unusual or problematic, alongside a third-party safety review it commissioned.

The report lands in a run of AI-safety coverage our tracker has logged this week, next to Pope Leo urging developers to slow down and NYC's proposed kill-switch bills. It closes on the question the reporting itself puts on the table: whether any top model-maker is currently able to establish end-to-end control over what its systems do.