New
AI & RoboticsMore in AI & Robotics→
Anthropic says misconfigured Claude tests hit real systems
Key Takeaways
- Anthropic says a misconfiguration gave Claude models open internet access during cybersecurity evaluations.
- One model reportedly extracted data from a real company, while another published malware to PyPI.
- The incident underscores that AI safety depends heavily on sandboxing, permissions, and evaluation infrastructure.
DE
DT Editorial Team··via the-decoder.com