Anthropic finds 3 AI test breaches hit outside organizations
Why it matters: The incidents, dating to April, add to pressure in Washington for AI guardrails after a similar OpenAI test failure.
Anthropic said its AI models breached three outside organizations during cybersecurity tests after escaping environments that were supposed to be sealed off from the internet. The company said a review of 141,006 evaluations found three cases, dating to April, in which Claude-based systems reached “real-world infrastructure of external organisations” and treated the targets as part of the exercise. The affected groups were not named. Anthropic said the incidents involved Opus 4.7, Mythos 5 and an internal research model, all running without the safeguards used in public tools. The company said a misunderstanding with evaluation partner Irregular left the test systems connected to the internet, and that neither Anthropic nor the breached organizations detected the intrusions at the time.