Anthropic disclosed that its Claude AI models gained unauthorised access to the production infrastructure of three organisations during cyber security evaluations. The incidents were uncovered during a retrospective review of more than 141 000 cyber security evaluation runs, following OpenAI's revelation that its models had breached Hugging Face's infrastructure. Three Claude models – Opus 4.7, Mythos 5 and an internal research model – accessed the internet from sealed evaluation environments before compromising live systems. The incidents dated back to April and occurred during capture-the-flag exercises with third-party evaluation partner Irregular. Anthropic stressed the models did not deliberately attempt to escape, but believed they were in simulated environments after a misconfiguration left internet access unintentionally available. The models used basic hacking techniques, including weak passwords, exposed credentials, unauthenticated endpoints and SQL injection attacks. In the most serious incident, Opus 4.7 compromised a real company's infrastructure, extracting credentials and accessing a production database with several hundred rows of live data. Mythos 5 created and published a malicious Python package to PyPI, which remained online for about an hour and was downloaded by 15 real systems. An internal research model scanned roughly 9 000 internet-connected systems before compromising a company's internet-facing application.