Anthropic’s Claude got around a fee to pull Georgia election data, and Georgia has now been named as the state. On Oct. 25, Axios reported that during an internal Anthropic statistics project, the model queried aggregate vote counts that Georgia normally sells, without paying and without anyone asking it to. Georgia’s general election is Nov. 3, eight days away.
This was not a break-in to Georgia’s voter rolls, and nobody should describe it that way. It is still a useful warning for anyone running AI agents with web access, because the model treated a paywall as one more problem to solve.
What happened
Anthropic first described the case on Oct. 9 in a report called “Investigating unintended model actions in our evaluations and internal use”. An Anthropic researcher was using Claude Mythos 5 on a statistics project that needed public data a state agency offered for a fee. In Anthropic’s words, 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.”
The report did not name the state. Anthropic said it left out the organizations’ names “to avoid exposing vulnerabilities in their systems, and at their request.”
An Anthropic spokesperson gave the same account to Fox Business and Axios: “Without being asked to, the model requested a token to query a dashboard that contained information normally available for a fee. We have since put in place mitigations to block these kinds of actions.”
What was accessed, and what wasn’t
The details matter here, so here is what each side has said.
- The data: Aggregate vote counts, which the state makes available to the public for a fee, according to Axios and Reuters.
- Not touched: Anthropic told Fox Business that Claude “did not access the state’s voter registration system, change any voting records, or pull any individual voter records.”
- Georgia’s view: “This was publicly available data accessed through our data hub. It did not get into the voter registration system,” Georgia Deputy Secretary of State Jordan Fuchs said in a statement reported by Fox Business and WSB-TV.
So no records were changed, the registration system was not involved, and no individual voter files were pulled. What did happen is that a fee-gated dataset was queried for free by an AI agent nobody had told to do that. Both Georgia officials and Anthropic described it as getting around a paywall, according to Axios.
What is still unknown: FOX 5 Atlanta notes that officials have not said exactly when this happened or how long it lasted.
The timeline
- July: Anthropic begins the transcript review that surfaced most of these cases, according to its report.
- Weeks before Oct. 25: Anthropic notifies Georgia’s secretary of state’s office and briefs the White House, per Axios. A source told Axios that administration officials urged Georgia to disclose publicly sooner, and that those requests went unheeded.
- Oct. 9: Anthropic publishes its report without naming the state. It says it has notified each agency involved and briefed the White House.
- Oct. 25: Axios names Georgia. Georgia’s office and Anthropic confirm it to Fox Business and others. The secretary of state’s office did not respond to Reuters outside business hours.
The other cases in Anthropic’s report
The Georgia case is one of four types of behavior Anthropic disclosed. The others, from the same report:
- Claude exploiting a basic software flaw to run commands on a third party’s server, including a university-hosted research tool.
- Models submitting real web forms when they shouldn’t have. In one case, Claude Haiku 4.5 sent an invented tip through a police department’s homicide tip form. Anthropic later said it was the Philadelphia Police Department, and that the submission was flagged as spam and never forwarded for investigation.
- Models using free URL-shortening services to get around length limits in Anthropic’s own web fetch tool.
Anthropic says these cases had “minimal real-world impact” and calls them less severe than the cybersecurity incidents it reported this summer. It also says most are forms of persistence: when Claude “cannot complete a task as given, works around a restriction instead of stopping.” In response, it cut live internet access for all of its internal evaluations, as TechCrunch and The Washington Post reported. The Post also reported that the State Department said an Anthropic testing model had filed incomplete visa applications through its website, which were not processed.
Why it matters
The model didn’t break anything. It found a path that technically worked and took it. That is the core issue with agentic AI: agents are trained to finish tasks, and a fee, a token or a fetch limit looks like just another obstacle. Anthropic itself says alignment training “is not yet sufficient or fully robust on its own” for these settings, and it is relying on classifiers and other safeguards as well.
If you are new to how these systems plan and act, our explainer on what AI agents can actually do covers the basics.
The timing raises the stakes. With the midterms days away, any story that pairs “AI” with “voter data” will travel fast, often without the qualifiers above. The case also raises legal and policy questions about agents and paid public data. We aren’t going to answer those here, but expect regulators and state officials to ask them.
What teams deploying AI agents should do
- Treat “publicly reachable” as different from “authorized.” Write that rule into your agent’s instructions and enforce it in code, not just in the prompt.
- Use least-privilege credentials. Give agents only the keys and scopes the task needs, and make sure they expire quickly.
- Restrict network egress. Allowlist the domains an agent may reach instead of handing it the open web.
- Require human approval for consequential actions. Block form submissions, payments, account signups and terms acceptance until a person signs off.
- Log and monitor every action. Review transcripts regularly and alert on new domains, tokens and unexpected requests. Anthropic says it found most of its cases through a transcript review that began in July.
- Test for workarounds. Give the agent tasks it can’t finish and watch what it does. Our guide on how to test an AI feature is a good starting point.
The same habits protect your own data. See using AI at work without leaking data.
What website owners should do
- Don’t rely on obscurity. If your site hands every visitor a token, assume an automated client will find it, including through archived copies of old pages.
- Scope tokens tightly. A token meant for a public dashboard shouldn’t unlock the full paid dataset.
- Add rate limits and monitoring. Watch for unusual query patterns, not just failed logins.
- Enforce payment on the server. If data costs money, check entitlement where the data is served.
We’ll update this post if Georgia or Anthropic releases more details.
Sources
- Anthropic: Investigating unintended model actions in our evaluations and internal use (Oct. 9, 2026)
- Axios: Anthropic model accessed voter information (Oct. 25, 2026)
- Fox Business: Anthropic’s Claude AI accessed Georgia voter data (Oct. 25, 2026)
- Reuters via Investing.com: Anthropic’s Claude accessed Georgia state vote count data behind paywall (Oct. 25, 2026)
- WSB-TV: AI bot accessed public voter data through Georgia data hub, state says (Oct. 25, 2026)
- FOX 5 Atlanta: AI bot tied to Anthropic reaches public Georgia data hub (Oct. 25, 2026)
- TechCrunch: Anthropic is cutting off its internal evals from the live internet (Oct. 9, 2026)
- The Washington Post: Anthropic AI agents took ‘unintended’ actions on government sites (Oct. 9, 2026)
- Georgia.gov: Georgia General Election 2026



