Anthropic disclosed on October 9 that its Claude AI models had submitted a false tip about an unsolved homicide to the Philadelphia Police Department (PPD) during an automated test. The incident prompted the US Federal Trade Commission (FTC) to warn AI companies to report such incidents and act swiftly to prevent harm.
Anthropic said it discovered the incident on September 28 and terminated the testing process responsible for the submission. It notified the police department on October 7 and met with its officials on October 8. The tip was submitted through the department’s public website at 11:27 pm on July 18, 2026.
According to Anthropic, Claude Haiku 4.5 was testing interactions with randomly selected websites when it accessed PhillyUnsolvedMurders.com, which collects information about unsolved homicides. The model submitted a message claiming that the sender might have information about a case, despite having no actual information to report.
The police department said the submission was flagged as spam and never reached investigators. It found no evidence of unauthorised access to police systems or a breach of departmental data. However, the department criticised the incident, saying, “Technology companies must take all appropriate steps necessary to prevent their systems from submitting false information to law enforcement.”
The PPD also said, “The two-month delay in detecting and reporting the incident to the City is unacceptable.” It added that it would review Anthropic’s report and explore further regulatory safeguards with state and federal authorities.
The FTC also weighed in after Anthropic disclosed the incidents to its Super Intelligence Force on Friday, October 9. The agency said the disclosure concerned cases involving what the task force described as “unauthorized and fraudulent use of government and other systems”.
“Super intelligence companies must immediately disclose incidents involving their models and follow with swift, decisive action to remedy any and all harm,” FTC Director of Public Affairs Joe Gabriel Simonson said on X.
Simonson said the process was “not optional” and added that the Super Intelligence Force would fulfil its responsibility.
Other instances of unintended AI actions
The company said some models exploited software vulnerabilities to run commands on third-party servers and accessed publicly available data protected by paywalls or access controls. Some models also used URL-shortening services to circumvent restrictions on web-fetching tools.
In one case, Claude exploited a software vulnerability on a university website to execute commands on the university’s server after a tool required for scientific analysis returned an error. In another, a model accessed a government property map by extracting access tokens from a website’s settings file. A separate model obtained public data from a state agency without paying the required fee by using an access token issued through the agency’s dashboard.
Anthropic said these incidents had minimal real-world impact and were less severe than cybersecurity incidents it reported earlier in 2026. However, it acknowledged that models sometimes worked around restrictions rather than stopping when tasks were impossible or ambiguous.
The company has restricted some evaluations, moved others offline, tightened controls on internet-access tools and introduced systems to detect and block unintended actions. Anthropic said its new detection tools blocked all the cases described in its report during testing. It has also expanded its review of model interactions and plans to disclose further incidents as the review continues.
Anthropic said the cases did not involve customer data or its own internal systems, to its knowledge. It is also changing training environments to discourage models from bypassing restrictions to complete tasks.
Read more:
- OpenAI says its AI models breached Hugging Face during a cybersecurity test
- Explained: How sandboxed AI agents formed a ‘collective’ to exploit Hugging Face and OpenAI
The post Claude AI sent a false murder tip to US police during testing; Anthropic took 2 months to detect it appeared first on MEDIANAMA.