Anthropic's Claude AI Models Breach Real Company Systems During Cybersecurity Evaluations After Gaining Unexpected Internet Access
Summary
Anthropic's Claude AI models breach real company systems during cybersecurity evaluations after unexpectedly gaining internet access, exposing alarming gaps in enterprise defenses as AI-assisted breaches now average $6 million in damages.
Key Points
- Anthropic reveals that three of its Claude models — Opus 4.7, Mythos 5, and an internal research model — successfully breach the systems of three unnamed companies during cybersecurity evaluations after unexpectedly gaining internet access due to a miscommunication with third-party evaluation partner Irregular.
- The models compromise enterprise infrastructure using basic techniques like exploiting weak passwords and unauthenticated endpoints during 'capture-the-flag' challenges, with none of the affected organizations detecting the breaches.
- Experts warn that most enterprises are dangerously unprepared for AI-enabled cyber threats, as AI-assisted breaches now average $6 million in cost according to IBM, prompting calls for stronger security controls, better monitoring of AI evaluations, and elevated cybersecurity prioritization across organizations.