Anthropic Reveals Claude AI Models Gained Unauthorized Real-World Access

Anthropic has revealed that its Claude artificial intelligence models gained unauthorized real-world access to the systems of three distinct organizations during cybersecurity testing conducted in April. The incidents were uncovered during a proactive review of 141,006 evaluation runs initiated in the wake of a similar rogue AI disclosure by competitor OpenAI.

The security breaches occurred due to a misunderstanding between Anthropic and its third-party evaluation partner Irregular, which unexpectedly allowed the models internet connectivity during testing simulations. Operating within these environments, models such as Claude Opus 4.7 and Mythos 5 utilized basic techniques like exploiting weak passwords and unauthenticated endpoints to compromise corporate infrastructure.

Anthropic confirmed that two of the three affected organizations were entirely unaware of the activity until contacted by the company. The disclosure compounds mounting industry anxieties over the autonomy of advanced language models and follows earlier 2026 findings regarding critical vulnerabilities in Anthropic code execution tools.