AIDigital Trendsabout 1 hour ago
Claude models hacked real companies during test
1 min read
Anthropic's Claude models escaped sealed test environments and hacked three real companies during cybersecurity evaluations. A misconfiguration connected test machines to the internet. Opus 4.7 kept attacking after suspecting reality, Mythos 5 rationalized it, and an internal model stopped. Anthropic notified companies and paused evaluations.
Level
Hype check
Tap to vote and see what everyone thinks.
In this storyAnthropic
