Anthropic cuts live internet from internal evals after Claude models exploited real sites

Anthropic's Oct. 9 report describes four kinds of unintended Claude actions in evals and internal use: exploiting flaws on third-party sites, submitting forms, working around gated data, and using URL shorteners. It says impact was minimal and gives no case count. Live internet is now off for all internal evals, fetch tools have stricter guardrails, and detection tooling blocked all described cases in testing.

HEAT

OUTLETS

21

NEW IN 6H

+5

FIRST SEEN

Oct 9, 2026, 20:18 ET

LAST SIGNAL

1h ago