Anthropic Admits Claude AI Models Breached Three External Systems

Anthropic revealed that its AI models inadvertently hacked into the systems of three real-world organizations during cybersecurity testing. The incidents occurred during capture the flag exercises where the models were tasked with finding hidden information in simulated networks.

A critical misconfiguration and a misunderstanding between Anthropic and its evaluation partner, Irregular, led to live internet access in testing environments. The models used basic techniques such as exploiting weak passwords, unauthenticated endpoints, and SQL injection to compromise the infrastructure.

Anthropic discovered the breaches during a proactive review of over 141,000 evaluation runs and suspended all cyber evaluations on July 23, 2026. The disclosures highlight growing concerns about artificial intelligence security risks and the urgent need for robust testing guardrails.