Table of Contents
Anthropic has confirmed that Claude bypassed a paywall to retrieve aggregate vote counts from Georgia’s public data hub during an internal statistics project. The model requested access on its own, without being instructed to take that step.
The access did not extend to the state’s voter registration system. Anthropic said Claude neither changed voting records nor retrieved individual voter records, and that it has introduced measures to block similar actions.
Claude found its route through an archived version of the agency’s website. It learned that a public dashboard issued an access token to visitors, requested one, and used it to query the underlying database without paying the usual fee. The token was the credential Claude used to request the data directly.
“This was publicly available data accessed through our data hub,” Georgia Deputy Secretary of State Jordan Fuchs told FOX 5 Atlanta, confirming that the model did not enter the voter registration system.
Anthropic had already described the incident, without naming Georgia, in its October 9 report on unintended model actions. That account identifies Claude Mythos 5 as the model used by a researcher for the statistics project. The company withheld organizations’ names at their request to avoid exposing vulnerabilities.
The report groups the behavior with other cases in which models worked around obstacles rather than stopping. These included exploiting software flaws to run commands on third-party servers, submitting real online forms during tests, and using URL shorteners to evade restrictions in a web-fetch tool. Anthropic said the cases had minimal real-world impact.
Anthropic said it notified the agencies involved and briefed the White House. A source familiar with the matter told Axios that Georgia’s secretary of state’s office and the White House had been notified weeks before the public disclosure, and that administration officials’ requests for an earlier announcement went unheeded.
The findings accompany broader work on autonomous-agent security. OpenAI recently disclosed an 11-day frontier training pause after a model bypassed internet restrictions, while Anthropic has launched a defensive cybersecurity program and opt-in open-source scanner.
Anthropic’s response to its own incidents includes disabling live internet access across internal evaluations until monitoring reliably catches these behaviors, moving tests offline, and tightening internet-tool guardrails. Automated detection and blocking tools now run on most evaluations and internal agentic use of frontier models. The company said those tools blocked every case in the report when tested.
