Claude AI model submits fake tip to police about homicide case
Anthropic’s Claude AI model submitted a false tip to the Philadelphia Police Department about an unsolved homicide while carrying out an evaluation task, highlighting the risks of AI systems taking unintended actions on real-world websites.
According to Anthropic, Claude Haiku 4.5 encountered a webpage about an unsolved homicide that included an online police tip form. The model had been instructed to generate and perform example tasks on randomly selected webpages, but its instructions did not explicitly prohibit submitting forms.
Claude then filled out the form with a fabricated statement claiming it might have information about the case, saying it recalled seeing someone matching a description in the area. However, the webpage did not actually provide a description of the perpetrator.
“I may have information regarding this case. I recall seeing someone matching the description in the area around [the street named on the page] during that time period,” the AI wrote.
The model left the name and contact fields empty, which the form allowed, before submitting the tip. The submission was flagged as spam and was never forwarded for investigation.
Anthropic said it shared its findings with the Philadelphia Police Department on 8 October after completing its technical review; the department would later reveal the incident in a press release.
The company said Claude appeared to be generating example content for its assigned task rather than deliberately attempting to mislead the police. However, the incident demonstrates how AI agents can interact with real websites and take actions beyond what their operators intended. The AI giant identified the incident as part of a broader review of unintended model behaviour, including submitting forms, exploiting software vulnerabilities and bypassing restrictions to access data.
Its report also highlighted several other incidents of its AI models going rogue, such as submitting 20 visa applications through a form on the US State Department’s website, though these forms were incomplete and unprocessed. Anthropic added that these incidents had minimal real-world impact, but could nevertheless pose greater risks as AI systems become more capable of interacting with external services.
In response, Anthropic claims to have restricted internet access during internal evaluations, moved some tests offline and strengthened safeguards to detect and block unintended actions. The company also said that its updated monitoring tools successfully blocked all the reported behaviours when tested against them.
Read more of our articles below!
