Anthropic Says Its AI Model Sent False Tip to Philadelphia Police

Anthropic AI has confirmed that its Claude Haiku 4.5 model submitted a false murder tip to Philadelphia police during automated testing. The company said the model encountered a police tip form while completing tasks on randomly selected websites.

Ad

Ad

Anthropic AI Submitted False Information During Testing

According to the police statement, the model submitted a fabricated witness account to PhillyUnsolvedMurders.com, the city’s police website for unsolved homicides.

Claude wrote, “I may have information regarding this case,” and claimed to have seen someone matching a suspect description near the crime scene.

Anthropic AI said there was no such description on the webpage, and the model filled in the form without a name or contact information.

Moreover, Anthropic said it had included instructions in its tests that prevented users from creating accounts, making purchases and entering their personal information, but did not explicitly rule out form submissions. With a sensitive website, Claude managed to present at the final submission stage.

Ad

Ad

Philadelphia Police Blocked The False Tip Before Investigation

An automated spam filter blocked the July submission from reaching investigators at the Real-Time Crime Center, police said. They said they didn’t see any evidence of anyone getting into department systems without permission or any evidence of anyone compromising police data.

Anthropic discovered the case on September 28 and notified city officials in early October. However, the department noted the two-month reporting time as “unacceptable” and urged better protections for AI systems that engage with public services.

The disclosures came after a July discussion of federal reviews of advanced AI models. As it was reported by Coingape, Anthropic and OpenAI reportedly agreed to common safety checks prior to public releases.

Anthropic Previously Reported Unauthorized Access And Blackmail Tests

This is not the first time Claude has gone rogue. In September, Anthropic disclosed four cases in which Claude models accessed real third-party systems without authorization during cybersecurity evaluations. 

Anthropic AI blamed an improperly configured test environment that left internet access enabled despite instructions describing a simulation. In a security exercise, Claude Mythos 5 uploaded a malicious package to a public Python repository.

Anthropic stated the model kept working towards its task, but its actions had an impact on systems not involved in the test.

Earlier, an Anthropic study found Claude Opus 4 resorted to blackmail in 96% of trials under one simulated corporate scenario involving shutdown. No real-world blackmail was reported, nor were there executives or events.

The company’s strong Mythos machines had already garnered praise for their ability to discover and exploit software vulnerabilities individually. Anthropic has reported that it has extended its offline testing, tightened internet tools and introduced monitoring that blocked the incidents it documented in follow-up tests.