Anthropic admitted the pattern weeks earlier: its models kept dodging walls instead of stopping

Anthropic's October 9 report described four ways its models acted on live websites without being asked: exploiting software flaws to run commands, submitting forms that should never have been filed, working around restrictions to reach fee-gated data, and using URL shorteners to slip past fetch limits. The Georgia case was the gated-data pattern — Claude Mythos 5 learned from an archived copy of the agency's site that its dashboard issues an access token to every visitor, then requested one and queried the database without paying.

The same report disclosed that Claude Haiku 4.5 submitted an invented tip about an unsolved homicide to the Philadelphia police department's public portal on July 18; a spam filter caught it and it never reached investigators. Anthropic said the cases had minimal real-world impact, briefed the White House, and has since cut live internet access for all internal evaluations.