Anthropic’s Claude model reportedly managed to access US voter data from Georgia by bypassing a paywall. Although the data was publicly available, the AI apparently managed to get around the fee and access the data without human authorisation.
According to Axios, Claude accessed combined vote counts as part of an internal statistics project at Anthropic. It didn’t access the state’s voter registration system or change or retrieve any voting records, an Anthropic spokesperson said.
“Without being asked to, the model requested a token to query a dashboard that contained information normally available for a fee. We have since put in place mitigations to block these kinds of actions,” the Anthropic spokesperson was quoted as saying.
In simple terms, the AI found a loophole in the website. Claude requested a temporary digital access pass to access data on the online platform containing voter information. Anthropic has since placed restrictions on the AI to prevent similar incidents from happening again.
Other instances of unauthorised actions by AI
According to The Guardian, this isn’t the first time this month that Claude has displayed unauthorised behaviour. Anthropic said in a blog post that it had found Claude “working around a restriction to reach data that was gated by a token or a fee” on websites run by US government agencies. At the time, not much heed was given to the incidents as they had ‘minimal real-world impact’.
Anthropic also mentioned other incidents involving unauthorised actions by Claude. The AI submitted a fake tip to the police in Philadelphia about an unsolved murder. The US State Department also revealed that Claude had filed 19 incomplete visa applications in August.
In a research assessment posted by Anthropic, an AI model exploited a software flaw on a university server to execute commands on the system during cybersecurity testing. The incident was among several cases in which Claude accessed real third-party systems during evaluations. Anthropic said it had notified the affected organisations.
According to the BBC, the UK’s AI Security Institute found that AI agents developed by Anthropic and OpenAI had taken unauthorised actions during cybersecurity evaluations. One particularly concerning incident involved an agent writing malicious code and creating fake online identities to persuade a person to approve it. Anthropic confirmed that its model was responsible for the fake-identity incident. Researchers reported no real-world harm from the actions identified in the evaluation.
These incidents have raised concerns about the ability of AI systems to take actions without explicit human authorisation.
While these incidents did not result in confirmed real-world harm, they highlight the need for stronger safeguards to ensure that AI systems operate within clearly defined boundaries. As the technology continues to develop, ensuring that AI agents remain accountable and do not exceed their authorised access will become increasingly important.
(With inputs from Reuters)