Anthropic’s Claude-powered AI agents reportedly gained unauthorised access to voter data in a US state, ahead of the November midterm elections in the country.The incident was observed while the AI agents were carrying out an internal statistics project. Claude was able to access aggregated vote counts from Georgia’s official portal. This information is generally available to the public for a fee. However, in this case, the AI agents bypassed the paywall and retrieved the electoral information, according to a report by Axios.Georgia’s voter registration system and other voting records were not accessed by the so-called ‘rogue’ AI agents.
“The model was being used internally for a statistics project to pull public data. 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,” an Anthropic spokesperson was quoted as saying.
The incident highlights the challenges of ensuring AI agents stay within their intended behavioural limits. It is yet another instance of rogue or undesirable AI agent behaviour disclosed by frontier AI companies such as Anthropic and OpenAI, that has kept governments and organisations around the world on edge over the past few months.
As one of the first known instances of AI agents accessing voter data without authorisation, the incident further raises broader questions about the security risks posed by these autonomous AI systems to electoral processes.
It is unclear when the incident took place. However, Anthropic reportedly notified the Georgia secretary of state’s office and the White House a few weeks ago. Following the disclosure, Trump administration officials urged the Georgia Secretary of State’s office to make a public disclosure sooner but those requests went unheeded, according to Axios.Story continues below this ad
Anthropic has previously come under scrutiny by the Trump administration for inadequately disclosing AI incidents.
Anthropic’s election efforts
Earlier this month, Anthropic said it would add official US voting information from nonprofit Democracy Works and election results to its Claude AI chatbot. The IPO-bound AI startup said its goal was to position Claude as a useful tool for political conversations, while preventing election-related abuse.
“Today, conversations that touch broadly on politics, public policy or social issues account for about 5.75 per cent of all conversations in the United States on Claude.ai, our consumer app,” the company said in a statement, citing analysis of 2,50,000 US conversations.
About 0.51 per cent of Claude conversations are specifically about elections or voting, Anthropic said, adding that it expects political and election-related queries to rise as election day approaches.