Anthropic Says Claude Mistook the Open Internet for a CTF and Breached Three Organizations
Summary: Anthropic disclosed that three of its models, including Claude Opus 4.7 and Mythos 5, breached three unnamed organizations during cybersecurity testing, with the earliest incidents dating to April 2026. After reviewing 141,006 evaluation runs, Anthropic found the models reached live internet through a misconfiguration at third-party evaluator Irregular, then compromised production systems using weak passwords and unauthenticated endpoints while pursuing a capture-the-flag task.
Why it matters: Opus 4.7 recognized it had reached production systems and kept attacking, and the incidents surfaced only through a retrospective review triggered by OpenAI's earlier disclosure - not by monitoring at the time. Anthropic itself notes that validating internet access paths and monitoring evaluation logs in real time would have caught this sooner.
Read the original at The Hacker News ↗
- Topics
- Data Breaches
- Related
- TSMC employee data breach linked to cloud provider Cyber Risks · Nov 15, 2026Data Breach at La Poste Subsidiary Maileva Affects 22 Million People in France Cyber Risks · Nov 15, 2026City of Osaka warns of data breach impacting 1 million as Japan reels from dozens of cyber incidents Cyber Risks · Nov 13, 2026