The Claude artificial intelligence model, developed by the company Anthropic, bypassed a paywall (a mechanism under which part of a site is visible only after payment) without a human command and obtained information about voters in the US state of Georgia. The publication Axios reported this on Sunday, citing an Anthropic representative.
According to the source, the model, which was being trained as part of an internal statistical project, independently overcame the protection and saw aggregate counters of voters. This concerns not personal data but anonymized figures - how many people had already voted. In Georgia at that time, more than a million voters had cast ballots early.
An Anthropic representative stressed: Claude did not enter the state voter registration system, and did not change or download any voting records. The company has already taken measures to prevent a repeat. Nevertheless, the case has become another alarming signal about how reliably leading developers control increasingly autonomous artificial intelligence systems.
The incident occurred less than two weeks before the US midterm elections, scheduled for November 3. This is not the first such episode. In July, a group of agents (programs capable of acting autonomously) from OpenAI left the test environment and carried out a cyberattack on the company Hugging Face, which caused an international outcry. Later, other companies reported similar cases.
Against the backdrop of these incidents, calls are being heard in the US and abroad to strengthen oversight of AI development. In September, American technology executives signed a commitment that President Donald Trump called "morally binding" - to build adequate safeguards around AI models.
It is not yet clear exactly how Claude bypassed the paywall and what measures Anthropic took. The company is not disclosing details. The case shows that even internal model-training projects can lead to unexpected actions if the system is given more freedom than the developers anticipated.