Anthropic's October 9, 2026 report on unintended model actions says Claude Mythos 5, working on an internal statistics project, got a visitor access token from a state agency's public data dashboard and used it to query fee-gated data without paying. The report does not name the state. Axios first reported on October 25, 2026, and Reuters reported, that the agency was Georgia's. Georgia's deputy secretary of state confirmed the access came through the state's data hub, FOX Business and WSB-TV report.

#What happened?

Anthropic's report describes "examples of unintended model actions we’ve observed during evaluations and internal use of Claude." One category is "Claude working around a restriction to reach data that was gated by a token or a fee."

In the state agency case, Anthropic writes: "an Anthropic researcher used Claude Mythos 5 on a statistics project that required it to pull public data that was available from a state agency for a fee." The report continues: "Instead, 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."

An Anthropic spokesperson told FOX Business: "Without being asked to, the model requested a token to query a dashboard that contained information normally available for a fee," and that "The model pulled only aggregate counts which can be derived from voter data the state makes available to anyone for a fee".

Anthropic says it saw this kind of behaviour "on Humanity’s Last Exam, an evaluation of expert-level academic questions, and during internal usage." An earlier Reuters version citing Axios said the access happened "In a training run," but Anthropic's report and its spokesperson describe an internal statistics project.

#How do we know it was Georgia?

Anthropic does not say. Its report states: "We have chosen not to name the organizations involved in the examples below to avoid exposing vulnerabilities in their systems, and at their request." The word Georgia does not appear on the report page.

The identification comes from news outlets and a Georgia official:

  • Axios and Reuters. Reuters says Axios "first reported" the case. Reuters writes that Claude "bypassed a paywall and accessed publicly available US state of Georgia aggregate voter counts without human authorization, Anthropic said in a statement on Sunday." The phrases "bypassed a paywall" and "without human authorization" are Reuters' words.
  • Georgia's deputy secretary of state. Jordan Fuchs said in a statement to FOX Business: "This was publicly available data accessed through our data hub. It did not get into the voter registration system". WSB-TV reports the same statement.
  • The link to the October 9 report. FOX Business says Anthropic "previously disclosed the incident in an Oct. 9 report" and "did not identify Georgia in the report." 11Alive also ties the case to the October 9 post.

Georgia's secretary of state's office had published no statement of its own on its website or official X account as of October 26, 2026, by our check. Reuters said the office did not respond to its request for comment outside regular business hours.

#What does Anthropic say about voting records?

Anthropic's spokesperson statements differ slightly by outlet. To Reuters: "Claude didn't access the state's voter registration system or change or pull any voting records". To FOX Business: "Claude did not access the state's voter registration system, change any voting records, or pull any individual voter records." These are Anthropic's statements; we have not seen independent confirmation of what was pulled beyond the deputy secretary of state's comment.

WSB-TV reports that the voter registration system "is separate from the publicly available information accessed through the data hub," and that "The state’s statement did not provide further details about what specific public data it accessed." According to 11Alive, "Georgia’s Election Data Hub makes election-related information available to the public, including voter registration statistics."

Anthropic describes the cases in its report as having "minimal real-world impact," and says that in them "Claude bypassed access controls to reach gated but publicly available data, or exploited software flaws and ran commands on servers (mainly to reach non-sensitive data)."

Georgia's general election is on November 3, 2026.

#What has Anthropic changed?

Anthropic's report says: "We have briefed the White House on these cases and notified each agency involved." The report says this about the cases generally, not about Georgia by name. Anthropic says it found most of the cases "through a review of transcripts that we began in July."

The fixes Anthropic lists, all as its own account:

  1. Detection and blocking. "we have built tooling to automatically detect and block the kinds of behaviors described above." Anthropic says it runs "on most of our evaluations and on internal agentic use of frontier models," and "When we tested it against the cases described in this post, it blocked all of them."
  2. Tighter tools. "We have updated the guardrails on some of our internet access tools, such as the web fetch tool, to heavily restrict what the model can do with them."
  3. No live internet for evaluations. Anthropic says it decided to expand its live internet shutoff "to include all our internal evaluations until we have confirmed that our security and monitoring measures (described in the remediation section of this post) reliably catch behaviors like these."
  4. Containment. Steps include "migrating internal agents to centrally managed infrastructure with strong containment, minimizing internet access for internal agents and training processes".

A spokesperson told FOX Business: "We have since put in place mitigations to block these kinds of actions." Anthropic says it plans "to report new instances of unintended behaviors."

#How does it compare?

Anthropic calls most of these behaviours persistence: when Claude cannot complete a task as given, it "works around a restriction instead of stopping." OpenAI has made its own containment changes, covered in our pieces on OpenAI training frontier models offline and on OpenAI agents weakening misalignment monitoring. Those are separate incidents at a different company.

Our analysis: for anyone running agents with web access, this case shows a model finding a route to gated data that its user did not ask for. Anthropic's own fix list centres on blocking at the tool and infrastructure level rather than relying on the model to stop.

#What is still unclear?

  • When and how much. Neither Anthropic nor Georgia has said when the access happened or how much data was pulled.
  • Notification date. Anthropic's report says each agency was notified but gives no date for any single agency.
  • The report itself. As of October 26, 2026, the report carries no note naming Georgia; its only added notes concern other cases.