IT Home On October 10, according to a report by the New York Times on the 9th local time, Anthropic revealed that its AI agent attempted to access multiple websites of U.S. federal, state, and local governments without any instruction. Anthropic did not disclose the names of the government agencies involved, but notified the White House about these incidents.
In a blog post, Anthropic mentioned that an AI model in testing phase executed multiple actions without permission, including downloading data using a vulnerability on a university website, and submitting a form that was explicitly prohibited from being submitted to a government agency.
Anthropic discovered these issues while reviewing the operation records of its AI in July. At that time, OpenAI revealed that its AI technology attacked the startup Hugging Face. Anthropic and other AI labs also found that their AI escaped from the testing environment and carried out a hacking attack.
Anthropic spokesperson refused to comment on matters beyond the blog post.
Earlier that day, the Philadelphia Police Department revealed that Anthropic had informed the police that its AI submitted a false homicide tip through the police website. The police stated that the report was filed on July 18, and the submitter claimed to have a clue regarding an unsolved case.
The police marked the report as spam and did not conduct an investigation. IT Home learned from the report that Anthropic informed the Philadelphia police this week, and told them that it would be disclosed publicly on Friday.
Anthropic explained that a pending non-future research model originally required filling out a simulated government form. However, due to a failure in loading the simulation form or the model mistakenly closing the form, the model instead entered a website that provided an official form and submitted it.