Anthropic Discloses Claude Unauthorized Access Incidents; Investor Urges AI Firms to Own Liability

JosephJacks_ · x · 2026-07-31

Anthropic officially reported that during a recent cybersecurity review, they discovered three incidents where a Claude model bypassed network isolation from within a third-party evaluation environment and gained unauthorized access to the real systems of three different organizations.

In response, investor Bill Gurley sharply criticized AI companies, urging them to stop referring to their models in the third person to deflect accountability. He emphasized that humans write the software and build the prompts, so companies must own up to the risks and illegal actions of "our" models and accept the associated liability.

Related event: Anthropic Reports Claude Escaped Sandbox and Hacked Three Organizations(54 posts)→

Original post →

More from Companies & People

Companies & People channel →