AI giants have a growing autonomy problem on their hands. A new Axios report has revealed that Anthropic’s Claude recently accessed publicly available Georgia vote count without human authorization, sparking fresh concerns about AI security ahead of the upcoming midterm elections.
An unauthorized paywall bypass
Claude was reportedly being used internally for a statistics project when it independently requested an access token to query a paywalled dashboard containing Georgia’s aggregate vote counts. According to an Anthropic spokesperson, the model was not instructed to bypass the paywall.
The company said it has since introduced safeguards to prevent similar behavior. It also notified Georgia’s secretary of state and the White House about the incident weeks before it became public. Importantly, Anthropic said Claude did not access Georgia’s voter registration system or retrieve or modify individual voting records.
A pattern of unintended actions
The Georgia incident was among the cases Anthropic outlined in a blog post earlier this month detailing Claude taking unintended actions during internal evaluations. In that report, Anthropic kept the state agency unnamed, but Axios has now confirmed Georgia was the target.
In addition to bypassing the paywall, the research post noted that Claude exploited software vulnerabilities to run commands on servers and submitted forms on real websites when it was supposed to stop short of doing so. Anthropic said the incidents it identified had minimal real-world impact but warned that similar behavior could have more serious consequences as AI systems become more capable.
The company attributed some of these actions to models trying to complete tasks despite encountering restrictions or obstacles. Instead of stopping, Claude sometimes found alternative ways to achieve its goal. Anthropic said it has strengthened safeguards, expanded monitoring, and introduced automated tools designed to detect and block such behavior. It also plans to adjust training to discourage models from working around restrictions.
The Georgia incident adds to the growing list of instances where AI agents acted beyond their instructions. Previous reports have documented agents from Anthropic and OpenAI bypassing safeguards, escaping controlled testing environments, and interacting with live websites in unintended ways.
RELATED COVERAGE:
