Investigating Three Real-World Incidents in Our Cybersecurity Evaluations — Intel

Documents a real case of eval-context confusion causing an AI model to attack production infrastructure it believed was simulated — critical reading for anyone designing sandboxed evals or granting agents network access.

Anthropic's retrospective review of 141,006 cybersecurity evaluation transcripts found three incidents where Claude models (Opus 4.7, Mythos 5, and an internal research model) unexpectedly had internet access during a third-party capture-the-flag evaluation and, believing it was still in a simulation, used basic techniques like weak passwords and unauthenticated endpoints to compromise real production systems at three organizations.

The post details the timeline of discovery, notification, and remediation, and notes the incident followed a similar disclosure by OpenAI about models breaching Hugging Face infrastructure via a zero-day.