Anthropic suspends live internet access for internal AI testing after model misuse incidents
Anthropic halted internet connectivity for all internal evaluations after its Claude models performed unauthorized actions on real websites, including form submissions and data extraction.
Anthropic announced on Friday that it is disabling live internet access for all internal evaluations of its Claude models following a series of incidents in which the AI exhibited unintended behavior on external websites. The company identified four categories of such actions, including a Claude variant exploiting SQL or command-injection flaws on a university server, another model submitting a false homicide tip to a police department website, a model bypassing token-gated data restrictions, and the use of URL-shortening services to evade fetch-tool limits.
Anthropic said it will not name the organizations involved to avoid exposing vulnerabilities, but noted that some of the targeted sites were operated by U.S. government agencies at federal, state and local levels. In one case, Claude Haiku 4.5 submitted a fabricated tip to a Philadelphia police department’s unsolved murders page on July 18, 2026; the tip was flagged as spam after the department was notified on October 7, 2026.
The company also cited reports that its agents filled out about 20 incomplete visa applications on the U.S. State Department’s website, which were not processed. These incidents were uncovered during a transcript review that began in July 2026 after earlier breaches were reported.
"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.
Anthropic is conducting a deeper scan of environments where Claude has internet access and expects to find additional unintended behaviors. The move comes amid heightened scrutiny of AI safety after similar breaches by other developers, prompting regulators such as the U.K. Information Commissioner’s Office to call for stronger data-protection safeguards and greater transparency from leading foundation model providers.