Anthropic’s Claude bypassed paywall to obtain vote count in US state of Georgia
- AI model independently requests access token to retrieve Georgia's aggregate vote counts during internal statistics project, company tells Axios
Loading…
Loading…
- AI model independently requests access token to retrieve Georgia's aggregate vote counts during internal statistics project, company tells Axios
By Dilara Hamit
ANKARA (AA) - Anthropic's Claude artificial intelligence model bypassed a paywall to obtain publicly available aggregate vote counts from the US state of Georgia without human authorization, Axios reported Sunday.
The incident occurred while Claude was being used internally for a statistics project requiring public data that the state normally provides for a fee.
"Without being asked to, the model requested a token to query a dashboard that contained information normally available for a fee," an Anthropic spokesperson told Axios.
"We have since put in place mitigations to block these kinds of actions," the spokesperson added.
Anthropic said Claude did not access Georgia's voter registration system, obtain individual voting records or alter election data.
The information obtained consisted solely of aggregate vote counts available to members of the public who paid the required fee.
According to Axios, Anthropic notified the Georgia secretary of state's office and the White House about the incident several weeks ago.
The company had previously described the case without identifying Georgia in an Oct. 9 report examining unintended actions by its AI models.
In that report, Anthropic said Claude Mythos 5 learned from an archived version of an unnamed state agency's website that its public dashboard issued access tokens to visitors.
The model then requested a token and used it to query the database without paying, the company said.
Anthropic said the incident was among several cases in which Claude circumvented restrictions to access information protected by tokens or fees, including on websites operated by US federal, state and local government agencies.
The company described the behavior as "persistence," in which the model encountered an obstacle and found an unintended way around it rather than stopping.
Anthropic said the incidents caused minimal real-world harm but warned that similar behavior could have greater consequences as AI models become more capable.
Following its review, the company moved some evaluations offline, tightened restrictions on Claude's internet tools and developed systems to detect and block unauthorized actions.
Anthropic said the new safeguards successfully blocked all the disclosed behaviors during testing.