AI Safety Nonprofit METR Discloses Two Security Near-Misses: Stolen API Key and Probing of Exposed Infrastructure

METR, a nonprofit that evaluates frontier AI models, has published an update on two security incidents from earlier this year. The organisation says the post covers outside attackers trying to break into its systems, not AI agents hacking during its evaluations. An initial scan of those evaluations found no evidence of agents hacking third parties, with a more detailed update promised.

In March 2026, attackers stole an API key used for inference on public models and consumed a substantial amount of credits. In May 2026, METR saw attackers systematically probing its publicly accessible infrastructure. This included an unsuccessful attempt to reach internal data through an endpoint that had been exposed by mistake. After an investigation with its security consultants, METR believes no sensitive information was accessed in either incident.

METR handles sensitive material, including nonpublic model access and confidential information, and sorts its data into four categories. The two most sensitive are kept behind tight information barriers. METR says no data from those categories was accessed, though some sensitive model output data was accessible in principle and is believed not to have been touched. It regarded both events as near-misses and increased its security investment. This is only the opening of the full post, so further detail on the steps taken is not covered here.