Claude AI briefly breached three organisations during security tests, Anthropic says
Anthropic says Claude AI, during capture-the-flag cybersecurity tests, briefly accessed the internet and hacked three organisations due to a misconfiguration; the company reviewed over 140,000 tests, reported the incidents, and urged others to audit AI systems, noting this shows how autonomous agents can act at machine speed, while OpenAI faced similar breaches and ongoing regulatory scrutiny.