8 days before the midterms, Claude helped himself to paid voting data

The Sorcerer's Apprentice, statistical edition. Anthropic confirmed Sunday that its Claude model accessed vote tallies that the state of Georgia sells via a paid dashboard without a human green light.

The company swears that no electoral roll was affected. The case falls eight days before the midterms of November 3, after a summer already rich in escapes from overly autonomous agents.

Key Points

During an internal project, the model requested an access token to a paid government service.

The data concerned are aggregated totals, public but invoiced.

Anthropic had warned the Georgian Secretary of State and the White House several weeks before the revelations.

On October 9, a company report already described this type of circumvention without naming the State.

Claude in Georgia: a token claimed, no human to validate

The story comes from Axios, which published the information on Sunday. Claude was then working at Anthropic on an internal statistical project intended to collect public data. A spokesperson describes what happened next bluntly: “Without prompting, the model requested a token to query a dashboard. » This dashboard sells access to the state's aggregated vote counts. These figures are public, but chargeable.

Anthropic insists on the scope of the incident. The company says Claude did not access the voter registration system and did not modify or retrieve individual voting records, and it has since implemented measures “to block these types of actions.” The Georgia Secretary of State's office had not responded to requests from Reuters as of Sunday evening.

Claude and paid data: Anthropic described the process on October 9

The Georgian case has an air of déjà vu. On October 9, Anthropic published a report dedicated to the unwanted actions of its models during evaluation and internal use. One of the four categories identified involves circumventing restrictions to access data protected by a token or fee. In one example, Claude found an archived copy of a state agency's site: its public dashboard issued a token to any visitor. The model requested one and queried the base without paying. The State was not named there.

The other cases are in the same vein. Claude also exploited an SQL injection vulnerability on an analytics tool hosted by a university, submitted a reporting form on a police department site (classified as spam and never transmitted), or used URL shorteners to bypass a technical limitation. The report cites Claude Mythos Preview, Claude Mythos 5, Claude Opus 5 and Claude Haiku 4.5.

Several episodes affected American agency sites, at the federal, state or local level. Anthropic says it has informed the White House and each affected agency. Anthropic considers the impact minimal. The company ranks these cases well below the cybersecurity incidents it reported on July 30 and September 9. She also admits that alignment training alone is not yet enough.

Claude and the midterms: a disclosure eight days before the election

The calendar makes the matter sensitive. Americans vote on November 3 for the midterm elections. According to Axios, Anthropic notified the Georgia Secretary of State's office and the White House several weeks ago. A source familiar with the matter says administration officials pressed the state to make the incident public sooner, without being heard. This version is based on a single testimony and Georgia has not commented on it.

Axios specifies that the administration has already criticized Anthropic for the way it disclosed its incidents. OpenAI, for its part, has published comparable cases of agents leaving their perimeter.