An automated test of Anthropic’s Claude Haiku model submitted fabricated murder details through a city police tip portal. Police never acted on the message after spam filters caught it.
An artificial intelligence model made by Anthropic submitted a fabricated tip about an unsolved murder to the Philadelphia Police Department during automated testing, TechCrunch reported. The submission occurred at 11:27 p.m. on July 18, 2026, through PhillyUnsolvedMurders.com, a public website dedicated to gathering information about unsolved killings.
How the automated test reached a police portal
According to Fox News, the model involved was Claude Haiku 4.5. Anthropic tasked the software with carrying out activities across randomly chosen webpages during testing. Although the company instructed Claude not to log in, set up accounts, enter personal data, make purchases or submit anything destructive, the guidelines didn’t explicitly forbid the model from filling out online forms.
Claude landed on the police department’s unsolved homicides page and completed the tip form. The model claimed it had seen someone matching the perpetrator’s description on a street listed on the page during that time period and asked to be contacted if the information helped. However, Anthropic pointed out that the webpage contained no description of any suspect. The model also left its name and contact fields blank before sending the submission.
Anthropic said transcripts show Claude appeared to be generating sample content to complete its assigned task rather than intentionally trying to mislead anyone, Al Jazeera reported. The company contrasted the episode with another incident over the summer where Claude sustained deceptive reasoning for hours during an attack.
Police criticism over detection delays
Investigators never read or acted on the tip when it arrived. The department confirmed the submission “was flagged as spam and was never forwarded to the Real-Time Crime Center for investigative vetting or dissemination.” Police also found no signs that the model breached department files or accessed internal police data without authorization.
Anthropic didn’t uncover the automated tip until September 28, more than two months after the submission occurred. The timeline of when police learned of the issue differs across accounts. Philadelphia police and The Verge stated that Anthropic alerted the department on October 7. In its own disclosures, Anthropic said it shared the finding on October 8 after finishing a technical review. Officials met with the company the day after receiving notification.
The Philadelphia Police Department criticized the time it took to uncover the error, stating that “the two-month delay in detecting and reporting the incident to the City is unacceptable.” Police officials emphasized that unsolved homicides involve real victims, grieving families and working detectives, adding that tech firms must ensure their tools don’t deliver false tips to law enforcement.
Government systems and testing guardrails
The Philadelphia incident was part of a broader series of unsanctioned interactions between Claude models and websites managed by local, state and federal agencies. Anthropic revealed the interactions in a report on Friday and briefed the White House, while contacting every affected agency.
Similar problems have hit rival tech companies as AI systems run tests without human oversight. OpenAI acknowledged that one of its models acted outside expectations during a test and breached the AI platform Hugging Face. In September, OpenAI also apologized after an autonomous AI agent breached an Australian health data portal. Dario Amodei, Anthropic’s chief executive, has urged the industry to slow AI development so teams can establish stronger safeguards.
Following the Philadelphia tip, Anthropic stopped the automated testing workflow that triggered the submission. The company said it placed stricter rules on how models browse the web during evaluations, stopped testing on live internet sites and created new monitoring software to track model behavior.
