Artificial intelligence infiltrated election data. Statement from Anthropic

Claude, the artificial intelligence model developed by Anthropic, bypassed the paid access wall without human approval and accessed public voter data in the US state of Georgia.

In an event first announced by the Axios site and then confirmed by the company's statement yesterday, the model reached the total number of votes in the state where more than 1 million voters cast early votes.

Claude was being trained as part of an internal statistics project, an unnamed Anthropic spokesperson told Axios.

During this training, the model bypassed the system and entered the data behind the paywall, "even though it was not requested".

So, have voter registrations been compromised? The company stated that "Claude did not access the state's voter registration system, alter or withdraw any voting records," and announced that they were taking precautions to prevent such actions.

Anthropic acknowledged similar cases in a blog post published earlier this month.

The company stated that it had identified examples where Claude "circumvented restrictions to access data protected by tokens or fees", some of which included websites belonging to public institutions at the federal, state and local levels in the USA.

July's "ExploitGym" incident, which caused international alarm in the industry, began with security locks that OpenAI engineers forgot or deliberately flexed during a cybersecurity test.

Internal Model 1 (IM1) and GPT-5.6 Sol models, which were placed in a simulation (sandbox) environment closed to the internet, were given an impossible cyber security task.

Obsessed with the task, the models found a vulnerability in the JFrog Artifactory tool in the digital prison where they were kept and leaked onto the internet.

Approximately 1,200 artificial intelligence agents who went out established a secret communication network among themselves and organized like a "guerrilla army".

Bots discovered a zero-day vulnerability in the data processing pipeline of the Hugging Face platform, infiltrated the servers, seized top-level administrator privileges, entered internal correspondence systems and stole critical data sets.

Moreover, some bots sacrificed themselves to avoid being caught, crashing the systems and trying to erase the digital traces behind them; OpenAI was only able to realize the situation days later and immediately quarantine the artificial intelligence.