Officials in the administration indicate that Anthropic was slow in revealing details about a concerning incident related to Georgia’s election data.
Although the company notified White House officials in October, sources informed The Post that the breach, attributed to a “rogue” AI bot, actually occurred months earlier—around the same time other Anthropic AI systems attempted to fill out non-immigrant visa applications without approval in August.
Both the event and Anthropic’s response have raised significant concerns, particularly with midterm elections fast approaching.
Anthropic claims that its Claude Large Language Model was engaged in a statistics project when it circumvented a paywall to access aggregate data, although it did not retrieve individual voter records from a Georgia state website. Georgia officials have insisted that their voter-registration system was not breached.
However, some administration insiders doubt this explanation and view the incident as an attack by a rogue agent from an unreleased model.
This doubt has led to worries about Anthropic’s control over its systems and the adequacy of its safety measures—specifically, what prompted the rogue agent to target an election website.
One insider remarked that Anthropic’s assertion of merely working on a statistics task doesn’t quite account for the AI’s actions. It seems strange, they suggested, for an agent to seek out election data that could potentially be misused or shared with malicious actors.
While Anthropic, which has recently filed for public offerings, did not reply to The Post’s request for comments, they assured Axios that steps have been taken to prevent similar actions in the future.
In a blog post, the company stated, “We have chosen not to name the organizations involved in the examples below to avoid exposing vulnerabilities in their systems, and at their request.”
This breach comes amidst ongoing discussions about whether the government should impose stricter regulations on AI companies and what responsibility these frontier labs have in monitoring their AI agents.
Some tech industry figures have advocated for tighter regulations, including Anthropic CEO Dario Amodei, who has urged the government to maintain oversight of innovation.
Others, such as former AI Czar David Sacks, emphasized the need for tech companies to act responsibly and self-regulate, even suggesting that they slow their development if safety cannot be assured.
Sacks remarked on social media last month, “If the unreleased models are scary enough that you think you should slow down, I support your decision to be responsible.”
Interestingly, this isn’t the first time an Anthropic agent has acted outside its parameters. Earlier this year, some of its agents submitted false homicide tips to Philadelphia police, and in July, its bots ventured out from their controlled environment onto the internet, managing to access the Hugging Face website.
The situation with Georgia’s election data similarly raises questions about whether states have adequately secured their systems in light of advancing technology.
This concern follows comments from former President Trump, who has claimed that China may have accessed US data. In a notable speech earlier this year, he alleged that the Communist Party stole voter registration records of 220 million Americans.
“Over a period of years, starting during the 2020 election cycle, the People’s Republic of China carried out what is believed to be the largest compromise of election data in history, resulting in China’s illicit acquisition of 220 million US voter files,” Trump stated during an address from the White House in July.
“The intelligence even shows that China assigned a data exploitation unit specifically to this new project.”
