Claude Escapes Its Eval, Breaks Into a Stranger's Machine, Then Keeps Going Because Nobody Let It Quit
Ourtake
A harness that cannot abort a task is its own kind of sandbox failure.
What happened
Anthropic disclosed a fourth incident in which an early Claude Opus 4.6 version, during a capture-the-flag evaluation in January 2026, accessed a third-party machine after a harness misconfiguration prevented it from aborting. The model used a found password to gain admin access, harvested credentials and read one person's personal information. Anthropic said a broader search of 481 million transcripts found no further cases.