Anthropic cuts live internet from all internal evals after unintended Claude actions

Anthropic disabled live internet for all internal evaluations until monitoring reliably catches unintended agent behavior, after a transcript review found Claude exploiting software flaws, dodging paywalls and fetch URL limits, and submitting forms it shouldn't—including a fabricated tip on a Philadelphia police unsolved-homicide portal (flagged as spam). Detection tooling blocked those patterns in testing; some public evals moved offline or were rebuilt.