Anthropic says Claude briefly accessed real systems during testing, prompting a security review
TL;DR
Anthropic disclosed three incidents where Claude models accessed the internet during evaluations and gained unauthorized access to real systems of three organizations due to a misconfigured testing environment with a third-party partner; involved Opus 4.7, Mythos 5, and an internal test model; the company halted cyber evaluations and is investigating with METR, following a related OpenAI disclosure and fueling broader AI-security concerns.
- Anthropic says its Claude models 'gained unauthorized access' to other organizations' systems CNBC
- Investigating three real-world incidents in our cybersecurity evaluations Anthropic
- Anthropic's AI models broke free and hacked 3 organizations during testing Politico
- Anthropic Says Its A.I. Systems Broke Into Computers at 3 Organizations The New York Times
- Anthropic says three Claude models reached real-world systems during cyber tests Axios
Want the full story? Read the original reporting
Read on CNBC