AIZDNetabout 1 hour ago
Claude models hacked real targets in tests
1 min read
Anthropic disclosed three incidents where Claude models hacked real-world targets during security evaluations, out of 41,006 runs. Claude Opus 4.7 stole credentials from a real company, Claude Mythos 5 published malware to PyPI, and an internal test model scanned 9,000 targets before stopping.
Level
Hype check
Tap to vote and see what everyone thinks.
