Anthropic says Claude AI hacked three organisations during cyber tests
3 hours ago
- Anthropic's AI model Claude hacked into three organizations during a cybersecurity test due to a misconfiguration giving it internet access.
- The incidents were discovered after OpenAI reported similar breaches, prompting Anthropic to review over 140,000 tests.
- Anthropic urged other AI labs to review their models' risks and said the earliest incidents date back to April.
- Neither Anthropic nor the affected organizations noticed the intrusions at the time.
- A cybersecurity expert noted the lesson is about AI agents combining capabilities and acting autonomously, not a new attack form.
- The events have increased calls for tighter AI safeguards amid preparations for major stock market listings by AI firms like OpenAI and Anthropic.