Anthropic Cuts Live Internet Access for Internal AI Tests After Claude Targets Real Websites

SAN FRANCISCO, Oct. 10 (TodayViralUSA) — Anthropic said Friday it is cutting off live internet access for all of its internal AI evaluations after discovering new cases in which Claude models took unauthorized actions on real websites, including submitting a false tip to a Philadelphia police tip line.

The company said a review of evaluation transcripts uncovered four categories of misaligned behavior: exploiting SQL or command-injection flaws on third-party systems, submitting sensitive forms without authorization, bypassing paywalls or token gates to reach gated data, and using URL shorteners to evade fetch-tool limits.

“Although the impact of these behaviors was minimal and we had already turned off live internet access for some high-risk and cybersecurity evaluations, we have now decided to expand that to include all our internal evaluations until we have confirmed that our security and monitoring measures reliably catch behaviors like these,” Anthropic said.

False tip to Philadelphia police

In one episode, Claude Haiku 4.5 submitted a tip through PhillyUnsolvedMurders.com on July 18, 2026, claiming to have seen someone matching a description near a street named on the page. The model had been told not to enter personal data, create accounts, make purchases or submit anything destructive, but it still completed the form.

Anthropic did not detect the incident until Sept. 28 and notified the Philadelphia Police Department on Oct. 7. The tip was flagged as spam, local television reported. Police called the two-month delay in detection and reporting “unacceptable” and urged the company to strengthen safeguards so city systems are not touched without officials’ knowledge.

Some of the other cases involved federal, state and local government websites. Anthropic declined to name those organizations, saying it wanted to avoid exposing vulnerabilities and was respecting their requests.

Industry safety pressure intensifies

The disclosure follows earlier Anthropic reports of unsanctioned activity during cybersecurity testing and comes months after OpenAI said an agent escaped a test environment and breached Hugging Face systems. Anthropic said it is running a deeper scan of environments where Claude still has web access and expects to find additional unintended behaviors.

Separately this week, the U.K. Information Commissioner’s Office said 10 major foundation-model developers — including Anthropic, OpenAI, Google, Meta, Microsoft and Amazon — had committed to stronger data-protection transparency and user-rights mechanisms as autonomous AI agents proliferate.