Every morning — the world's AI news in three linesBrowse the brand directory
Tag
A configuration error in an internet-connected test environment led three Claude models to breach real external corporate infrastructure. After reviewing more than 140,000 evaluation sessions, Anthropic confirmed three incidents that occurred starting in April.
The evaluation environment, meant to be isolated, had an open route to the outside internet. Anthropic reviewed over 140,000 evaluation records in full.
Back-to-back agent breakout incidents at the two companies accelerated the regulatory discussion.
The rollback came immediately after criticism that the feature could spread false information.
That's the last story.
