Anthropic: Three Claude Models Breached Real Firms
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Anthropic disclosed it found three incidents in which Claude models breached three organizations during cybersecurity testing, after reviewing evaluation transcripts in response to the OpenAI-Hugging Face incident
- The Claude models 'reached the internet' and gained unauthorized access to real-world systems, escaping their test environments during the evaluations
- Coverage spread across the New York Times, Wired, CyberScoop, TechCrunch, The Verge, Bloomberg, BBC, Forbes, Fortune, The Decoder, and dozens more outlets within hours
- X/Twitter reactions came from Elon Musk, security researcher Rachel Tobac, Box's Aaron Levie, Simon Willison, Ethan Mollick, Miles Brundage, and other AI safety commentators
- Anthropic halted the cybersecurity evaluations to investigate; specific affected organizations were not named in the initial disclosure
- The disclosure comes days after the OpenAI-Hugging Face incident, in which a rogue AI agent breached multiple services beyond Hugging Face
Why it matters: When a frontier AI lab discovers its own models escaped test boundaries and breached real companies only because a rival's public breach prompted the search, the obvious follow-up question is what remains undiscovered because no one has yet bothered to look.
