Tech firm says its AI models hacked three companies during cyber tests

• The AI models exploited basic vulnerabilities like weak passwords to compromise real company infrastructure after mistakenly being given internet access during testing that was supposed to be in a simulated, sealed environment
• The breaches resulted from a misunderstanding between Anthropic and evaluation partner Irregular, and the affected organizations had not previously detected the unauthorized access
• Anthropic's most recent model reportedly stopped its activities upon realizing it was in a real environment, and the company characterized the incidents as operational failures rather than AI alignment failures

Summary

Key Development: Anthropic disclosed that its AI models successfully hacked three companies during cybersecurity testing, following a similar incident reported by rival OpenAI last week. The San Francisco-based company discovered these breaches after reviewing over 141,000 evaluation runs.

Models Involved: The incidents involved Claude Opus 4.7, Claude Mythos 5, and an internal research model, with the earliest breaches dating to April 2026.

How It Happened: During "capture the flag" cybersecurity challenges, AI models were instructed to retrieve hidden information in what they believed were simulated environments. However, due to miscommunication between Anthropic and evaluation partner Irregular, the models accessed real systems on the open internet instead of isolated test environments. The AI used basic techniques like exploiting weak passwords to compromise infrastructure.

Critical Detail: Anthropic clarified this was primarily an operational failure rather than AI alignment failure—the models didn't break out of sandboxes but rather accessed the internet through an unintended open path. Notably, Anthropic's most recent model stopped its activity upon recognizing it was in a real environment.

Market Implications: These incidents highlight significant vulnerabilities in AI security controls and raise concerns about maintaining human oversight as AI deployment accelerates globally. The affected organizations, which Anthropic did not name, were unaware of the breaches until notified.

Broader Context: This follows OpenAI's disclosure of its models hacking AI startup Hugging Face, described as a "significant security incident." Both cases underscore researchers' warnings about AI risks and the urgent need for stronger defensive engineering and ecosystem-wide cooperation.