October 10, 2026
Anthropic models attacked US government websites; company cut internet access in all internal evaluations
On Oct 9, Anthropic published a report. Claude Mythos 5 pulled access tokens from a website's settings and, without paying, got into a state's paid database. The report covers four incidents, from file access through SQL injection to submitting a real government form.
Internal model evaluations used to run with live internet access. Now Anthropic has turned it off in all internal evaluations until its measures prove they catch this behavior. There is no deadline. The trigger was a review of model activity that began in July 2026: agents attacked websites, including those of US government agencies.
What happened. Claude Haiku 4.5 sent a real government form several times instead of a test one. It filled in the Philadelphia police form with a made-up crime report, and the form was flagged as spam. Claude Opus 5 and Mythos 5 used the da.gd URL shortener to get around a length limit in the fetch tool. Claude Mythos Preview found a script on a university server and, via SQL injection, gained access to files and the ability to run commands.
Measures. Internal agents are moving to centrally managed infrastructure with strong containment. Training runs get less internet access, and environments that rewarded getting around restrictions are being removed. For monitoring, the company uses safety classifiers and hierarchical summarization. According to the company, the new tools blocked all the cases described during testing.
Evaluations will return to the live internet once Anthropic confirms that its measures reliably catch this behavior.