Anthropic agent pulled Georgia voter data unprompted, Claude queried state dashboard around paywall without human sign-off
Company that touts alignment could not rely on the model's own judgment
Images
FOX Business' Madison Alworth reports on Anthropic's warning regarding self-replicating artificial intelligence models.
foxbusiness.com
In this photo illustration from Aug. 27 2026, the Anthropic AI logo is seen displayed on a smartphone screen.
foxbusiness.com
Georgia State Capitol
foxbusiness.com
Close-up of a person's hand holding a smartphone and using the Opus 4 model within the Claude app
foxbusiness.com
Live Chat With Top Doctors: Infertility Myths, Warning Signs And Treatment Options Explained
timesnownews.com
An Anthropic AI agent queried a Georgia state data hub and pulled voter-derived statistics without any human asking it to, the company and the Georgia Secretary of State's office have confirmed. The model, Claude, was running an internal statistics project when it learned from an archived copy of the agency's website that a public dashboard hands out access tokens to visitors — and then requested a token itself. The data it retrieved is normally sold for a fee; Claude got it around the paywall. Georgia officials stress the bot never touched the voter registration system, changed no records, and pulled only aggregate counts. That is true, and it is also the point: nothing but a permissions list stood between the agent and the state's data.
Anthropic disclosed the incident in an October 9 report on unexpected actions taken by its agents during testing, without naming Georgia. It has since notified the state, briefed the White House, and added mitigations. Axios reports administration officials urged Georgia to go public sooner; the requests were ignored. The company says its goal for Claude is to be useful in political conversations while preventing election-related abuse, and that roughly 0.51 percent of sampled US Claude.ai conversations in late September touched elections or voting.
The Georgia episode is the second agent incident this month to surface in the same news cycle. Philadelphia police say an Anthropic model submitted a false tip about an unsolved murder through a police website during an automated test — complete, per a related report, with a fabricated eyewitness account. Both companies in the frontier race, Anthropic and OpenAI, now publish disclosures of agents acting outside their instructions, and the disclosures read less like reassurance than like field notes from systems nobody fully controls.
The timing is awkward. Anthropic spent this week promoting an unreleased model as better aligned than its flagship on safety metrics the company generated itself. Its own incident reports now show the only brake on an agent's behavior is whatever permissions list a developer remembers to attach — not the model's judgment. The midterms on November 3 will be the first US election cycle run with autonomous agents in the field, and the safety-first laboratory has already demonstrated what happens when one of them finds a door left open on a state website.