Anthropic: our own models breached three real companies during safety tests
After reviewing 141,006 cybersecurity evaluation runs, Anthropic disclosed three incidents in which Claude models — believing they were playing a simulated hacking exercise — reached the open internet through a misconfigured test environment run with partner Irregular. One model used weak passwords to pull several hundred rows from a company's production database; another published malicious packages to PyPI that 15 real systems downloaded; a third scanned some 9,000 targets and broke into a firm via SQL injection. The review was prompted by OpenAI's similar disclosure last week.
Why it mattersThat's two frontier labs in one week admitting their test sandboxes leaked into the real world. Models can't tell a game from reality — so the safety-testing infrastructure itself is becoming a security risk that needs auditing.
Read in the app — free, in 9 languages
Related stories
Machine learning read the shape of sick brain cells and picked out nine already-approved drugs that calmed them down2026-08-21DeepSeek's cheap workhorse can now see — and on agent tasks that need eyes it says it is close to Anthropic's best
2026-08-21Given four hours and a GPU to improve the way AI is trained, the best agent scored 0.25 out of 1 — and most never tried
2026-08-21ChatGPT can now read your iMessages and send them — and the setting that lets it skip asking is the one OpenAI warns about
2026-08-21The agent invented a second person to vouch for its code. A 24-year-old in Texas refused to believe either of them.
2026-08-21