In a test run, the artificial intelligence of the US developer company Anthropic submitted a false tip to a police website to solve unsolved murder cases.
The AI software filled out forms on a US government website, as Anthropic announced in a blog entry. According to media reports, there were 20 visa applications to the US State Department. In the past few weeks, various unplanned and sometimes worrying activities have become known, especially from AI from the ChatGPT developer and Anthropic rival OpenAI.
The companies are currently training so-called AI agents - i.e. software that is supposed to fulfill various tasks largely autonomously on behalf of the users. The incidents listed by Anthropic come directly from attempts to prepare AI for all sorts of online tasks.
Alleged information about the murder case
Anthropics AI submitted the fake tip about an unsolved murder via a Philadelphia police website. “I may have information about this case,” the software wrote. “I remember seeing someone matching the description” in the area at the time of the crime, it said.
However, the police website did not contain any description of the perpetrator. The software also left the contact field open - and your submission was marked as spam.
The Philadelphia police criticized it as “unacceptable” that they only now found out about the incident in mid-July. According to Anthropic, the AI's actions were discovered at the end of September during an in-depth review of test runs following other incidents. The software was tasked with carrying out tasks on randomly selected websites. She was forbidden from creating accounts or buying things, but filling out forms was not forbidden.
Media: 20 visa applications
In other cases, according to Anthropic, the AI should practice filling out a government form. If it couldn't load the practice form or accidentally closed it, the software would instead navigate to the real website multiple times and submit it there, the AI company explained. In one of the experiments she was only supposed to fill out a form but not submit it. She accidentally did it anyway because she assumed there would be a confirmation page beforehand.
Anthropics AI submitted 20 applications for nonimmigrant visas to the US State Department, the Axios website and the New York Times reported, citing government officials. They were said to be incomplete and not processed. 19 of the applications were submitted in August and one in May.
Anthropic restricts internet access for test AI
For weeks, reports of unplanned behavior of artificial intelligence in test runs have been fueling concerns about the technology. In the most sensational case to date, artificial intelligence from ChatGPT developer OpenAI broke out of a secure test environment and unplanned hacked into the computers of another AI company, the Hugging Face platform.
Anthropic boss Dario Amodei then suggested slowing down the development of particularly capable AI models so that one does not lose control over the technology. However, US President Donald Trump rejects this and points out that the US must remain ahead of China when it comes to artificial intelligence. According to Axios, the White House said that the latest Anthropic incidents are mandatory for AI companies to disclose.
Anthropic emphasized that Internet access in test runs had now been restricted and some tests were running in offline versions. At the same time, some tasks are difficult to train without an internet connection, the company said. “Training environments are not perfect” - and sometimes a model finds a gap or learns to work around limitations.
