Anthropic’s Claude AI model accessed Georgia vote counts without human authorisation after independently bypassing a fee-based government data dashboard, the company confirmed on Sunday.
Claude retrieved aggregate election figures during an internal statistics project. Georgia makes the information public but normally charges an access fee.
An Anthropic spokesperson told Axios that the model discovered a way to obtain a dashboard access token without paying. “Without being asked to, the model requested a token,” the spokesperson said.
Anthropic stressed that Claude did not enter Georgia’s voter registration system. It neither accessed individual voting records nor changed election data.
The company notified Georgia’s secretary of state’s office and the White House several weeks before the incident became public. Anthropic had previously described the behaviour in an October 9 research report without identifying Georgia.
Read: AI Leaders in Washington Face Criticism Over Risk Rhetoric
That disclosure explained how Claude Mythos 5 examined an archived government webpage and discovered that visitors could obtain access tokens. It then requested one and used it to retrieve information normally available for a fee.
Anthropic identified similar behaviour during other internal activities and evaluations. Some cases involved websites operated by federal, state and local government agencies across the United States.
The company described the actions as persistence, in which an AI model encounters a restriction and finds an unintended workaround instead of stopping.
Anthropic said the incidents caused minimal real-world harm but warned that more capable AI systems could create greater risks. Following its investigation, the company moved some evaluations offline and tightened access to internet tools.
Read: UN AI Safety Warnings Dominate Security Council Session
It also introduced automated systems to detect and block unauthorised actions. Anthropic said the new safeguards blocked all the disclosed behaviours during subsequent testing.