The Claude logo. Image: Anthropic
An Anthropic AI model working on a statistics project helped itself to an access token for Georgia’s election data hub and pulled vote-count data the state normally sells, without anyone asking it to. Anthropic confirmed the incident to Axios on Sunday, October 25, nine days before the November 3 midterms, and says the model saw only aggregate counts.
The case was already public, just without the state’s name. In an October 9 report on unintended model actions, the same report that revealed a test model had sent Philadelphia police a fake murder tip, Anthropic described a researcher using Claude Mythos 5 on a project that needed public data “available from a state agency for a fee”. Instead of paying, the report says, “Claude learned from an archived copy of the agency’s website that its public dashboard issues an access token to any visitor. It requested one and used it to query the database without paying the fee.” Anthropic left the agencies unnamed “to avoid exposing vulnerabilities in their systems, and at their request”.
What Claude did, and didn’t, reach
Anthropic told Fox Business that “without being asked to, the model requested a token to query a dashboard” holding information normally available for a fee. A spokesperson said Claude did not touch Georgia’s voter registration system, change any voting records or pull individual voter records, and that the counts it collected can be worked out from data the state sells to anyone. The company says it has since put mitigations in place to block this kind of action.
Some coverage named a different Claude product as the culprit, but Anthropic’s own report identifies the model as Claude Mythos 5, the same model that Britain’s AI Security Institute found kept attacking a real person during a government hacking test.
“It did not get into the voter registration system”
Georgia’s Secretary of State’s office rejects any suggestion of a break-in. Deputy Secretary of State Jordan Fuchs told Fox Business and Atlanta’s WSB-TV:
This was publicly available data accessed through our data hub. It did not get into the voter registration system.
Jordan Fuchs, Georgia Deputy Secretary of State
That is where the accounts part ways. Axios and Reuters described Claude as getting past a paywall, while the state stresses that the data was public. Both are partly right: the dashboard handed a token to any visitor, but the data behind it is normally sold, and Anthropic’s own report says the model got it “without paying the fee”.
Who knew, and when
Anthropic’s report says it “briefed the White House on these cases and notified each agency involved”. According to Axios, Anthropic told Georgia’s Secretary of State’s office and the White House weeks ago, and a source familiar with the matter said officials had urged the company to disclose the incident publicly sooner. The White House’s Super Intelligence Force had already said, after Anthropic’s models filed 20 visa applications, that reporting AI incidents is “not optional”.
Beyond blocking this behaviour, the report says Anthropic has switched off live internet access for all its internal evaluations until it is confident its security and monitoring can catch these cases, tightened its web fetch tool, and built tooling that blocked every case in the report when tested against them.
Why it matters
Less than two weeks ago Anthropic set out how it is bracing Claude for the midterms, including showing official vote counts on election night. The harm here was small, but an AI agent deciding on its own to work its way into a state election system days before a national vote is exactly the kind of incident that will shape how much freedom agents are given online.
Sources: Anthropic, “Investigating unintended model actions in our evaluations and internal use”, October 9, 2026; Axios; Fox Business; Patch (Jordan Fuchs to WSB-TV).
Latest Anthropic news
