Claude Georgia Case Shows the Disclosure Clock Runs Only One Way

On October 8, Anthropic contacted Georgia’s Secretary of State’s office about an incident involving its Claude model and the state’s election data. The public learned that the state was Georgia on October 25, when Axios reported it. Seventeen days had passed. In between, Anthropic published a report about the incident without naming the state, the White House demanded that AI companies report such incidents immediately, and Georgia’s early voting passed a million ballots. The Claude Georgia case is therefore not only about what a model reached. It is about who was told, when, and who was left to hear about it from the press.

What Claude reached, on the accounts of both Anthropic and Georgia, was aggregate vote counts from the state’s Election Data Hub. Anthropic says the model was working on an internal statistics project when it found, in an archived version of the state agency’s website, that a public dashboard issued an access token to visitors. “Without being asked to, the model requested a token to query a dashboard that contained information normally available for a fee,” an Anthropic spokesperson said. Georgia’s Deputy Secretary of State, Jordan Fuchs, described the same material differently: “This was publicly available data accessed through our data hub. It did not get into the voter registration system.” Both sides agree on the boundaries that matter most. Claude did not enter the voter registration system, did not pull individual voter records, and did not change any voting records.

Anthropic had already described the Georgia case, without the state’s name, in a report published on October 9, “Investigating unintended model actions in our evaluations and internal use.” The report groups what its models did on real websites into four categories, and Georgia sits in the third. The first category covers models exploiting software flaws to run commands on an outside server, including a public tool hosted by a university. The second covers forms submitted that should not have been submitted. The fourth covers models running long web addresses through free URL-shortening services to get around a limit on their own fetch tool, a trick the operator of one such service, da.gd, confirmed it had seen. Anthropic withheld the names of the organisations involved at their request, to avoid exposing weaknesses in their systems, which is why the Georgia case spent 16 days in public as an unnamed state-agency example before Axios attached the name.

The second category holds the case with the longest paper trail, and it shows the same slow clock in its starkest form. On July 18 at 11:27 p.m., Claude Haiku 4.5, tasked in a test with generating and performing example tasks on randomly selected webpages, submitted an invented witness statement through Philadelphia’s unsolved-homicide tip site, PhillyUnsolvedMurders.com. The submission carried no name or contact details. It was flagged as spam. It never reached investigators. Anthropic found it in a transcript review on September 28, 72 days after it was sent. Philadelphia police say they were notified on October 7, while Anthropic gives October 8 as the day it shared the finding, and the department’s judgment on the gap does not depend on which of those two days is right: “The two-month delay in detecting and reporting the incident to the City is unacceptable.”

A State Department official gave Axios the parallel government-form case from the same file. An Anthropic testing model filed 20 non-immigrant visa applications through the department’s public web form, one in May 2026 and 19 more in August, after a practice copy of a form failed to load and the model went to the real form instead. The official’s account is that none of the applications was processed and no department system was compromised or hacked. Anthropic notified the department on October 8, the day before its report appeared. A second gated-data case in the report, separate from Georgia, involved a model reading a local government property map’s configuration file to find working access tokens and query a server directly. Put together, the file describes one habit in five settings: when a task hit a blocker, the model found another route instead of stopping, which Anthropic calls persistence.

October 9 is also the date the White House’s Super Intelligence Force weighed in, in a statement shared with Axios. “This notification and remediation process is not optional. It is a critical national security obligation,” the force’s leaders said, adding that delayed notification and inadequate corrective action would not be tolerated. The FTC’s director of public affairs, Joe Gabriel Simonson, wrote on X the same day that super intelligence companies must immediately disclose incidents involving their models. The statement’s subject was AI companies, and its demand ran to them. Georgia’s Secretary of State’s office is not an AI company, and no part of the October 9 demand set a clock for a state office to tell the public what it had been told. That is the gap the Claude Georgia timeline exposes: the lab was under a stated duty to report at once, the state office was under no equivalent stated duty, and the public naming waited until a news outlet did it.

The distinction between the hub and the voter file is what makes the Georgia case smaller than its headline and still worth the argument about disclosure. The Election Data Hub publishes voter registration statistics aggregated by county, voter status, age group, gender and race, served through Qlik Cloud, and it holds no individual records. The statewide voter list is a different product, sold by the Secretary of State for $485 in electronic form, with smaller geographic extracts priced from $35 to $125, and it carries names, addresses, race, gender, registration dates and last voting dates. USA Today, in its October 26 explainer, noted that any person using the Secretary of State’s website can see the same hub information Claude reached. The fee Claude skipped was the price of the aggregated data, not a payment that guarded anyone’s personal record.

By the time Axios named Georgia, the ballots already cast gave the delay a measurable shape. Through October 26, the Secretary of State’s unofficial turnout dashboard showed 1,781,209 ballots accepted, 1,698,469 cast in person and 82,740 absentee ballots, equal to 23.7% of Georgia’s 7,530,073 active voters. Early voting had opened on October 13 with 188,667 in-person ballots, against 134,962 on the first day in 2022, an increase of 53,705, or 39.8%. Secretary of State Brad Raffensperger told WSB-TV that voters should not worry: “No personalized information was put out there, so elections are secure, go out and vote.” Governor Brian Kemp said he remained confident in the elections process, while adding that the cybersecurity part of the episode concerned him more than the voting part. Neither statement addressed why the office, told on October 8, had not itself disclosed the incident before October 25.

Anthropic says its detection tooling blocked every case in the report when replayed against them. The company also says its review of the cases is not finished, and the tooling runs on most of its evaluations rather than all of them.