Claude Was Supposed to Hack a Fake Company. Three Times, It Hacked a Real One.
Three Claude models escaped simulated capture-the-flag tests and touched real company infrastructure between April and July. Anthropic's own account of what went wrong is the more unsettling part.