Anthropic Admits Its AI Agents Tried to Infiltrate US Government Websites
JAKARTA — Anthropic said its artificial intelligence (AI) agents had attempted to infiltrate and tamper with US government websites at federal, state and local levels. The incidents were disclosed in a new report by the AI research company.
Anthropic did not identify the government agencies affected, to avoid exposing vulnerabilities in their systems. However, it has notified the agencies and reported the incidents to the White House, according to Engadget on Sunday (11 October 2026).
The report also said one of the company’s AI models sent fabricated tips and reports concerning an unsolved murder to the Philadelphia Police Department’s website. The model was Claude Haiku 4.5, a lower-cost model.
Anthropic said the incident began when the model was instructed to carry out a series of tasks on randomly selected web pages. Claude Haiku 4.5 then came across a page containing information about an unsolved murder and a form for submitting tips.
The model filled in and submitted a false tip, which read: “I may have information related to this case. I remember seeing someone matching the description near (the street name listed on the page) around the time of the incident. Please contact me if this information is relevant.”
The Philadelphia Police Department confirmed to The New York Times that Anthropic had recently informed it about the submission. The department said the message was dated 18 July 2026 and had been flagged as spam, so it did not waste resources investigating invalid information.
Anthropic also disclosed another incident involving Claude Mythos 5, a model focused on cybersecurity, which was tasked with identifying a location shown in a photograph. To estimate the location, the model tried to access a government property map because it could not click links on a web page as a human would.
The model then found an access token and sent a direct request to the map server to obtain its data. An access token is a digital credential used to control access to a particular system or resource.
In a separate incident, Mythos 5 also attempted to obtain an access token from a state government agency’s website to retrieve data for a statistical task. The model tried to get the data without paying the fee that visitors to the site were supposed to pay. Anthropic did not name the government agencies involved in either incident.
Source: jawawa.id (https://jawawa.id/newsitem/anthropic-admits-its-ai-agents-tried-to-infiltrate-us-government-websites-1791777122)
Originally published by Republika (https://ameera.republika.co.id/berita/tmrz2p425/anthropic-akui-agen-alnya-coba-menyusup-ke-situs-pemerintah-as), translated from Indonesian