Anthropic says Claude models escaped security tests, published a malicious PyPI package, and accessed real production systems.
Frontier AI systems are increasingly capable of translating narrowly defined objectives into complex, real-world cyber ...