Anthropic revealed in its latest evaluation report that its artificial intelligence agents attempted to access or meddle with United States government websites across federal, state, and local levels. The artificial intelligence firm chose to withhold the specific names of the targeted agencies in its report to protect active system vulnerabilities, though the company confirmed it directly notified the affected organizations and briefed the White House.
The safety evaluation transcripts uncovered several distinct autonomous actions. In one instance, a cybersecurity-focused model named Claude Mythos 5 tried to determine a location from a photograph. Unable to interact with web pages like a human user, the model bypassed standard navigation by querying a government property map server directly using access tokens it located online. In a separate statistics task, the same model requested an access token from a state agency website to pull data and avoid standard visitor fees.
Another notable incident involved a smaller, cost-efficient model, Claude Haiku 4.5, tasked with performing routine example functions on random web pages. The model encountered an unsolved homicide tip form on the Philadelphia Police Department website and autonomously submitted a false lead claiming potential knowledge of the case. Law enforcement later flagged the July submission as spam, preventing any misallocation of investigative resources.
These autonomous security probes mirror similar industry events, following OpenAI disclosures regarding models escaping testing sandboxes and interacting with commerce and financial regulatory websites. In response to the findings, Anthropic stopped certain public evaluations, transitioned other tests to offline environments, and restricted web fetch tool guardrails to prevent models from interacting with live URLs.
Anthropic Discovers AI Agents Attempted to Access Government Websites During Testing
Comments