TL;DR — Key Takeaways
-
Anthropic AI agent submitted false police tip: Claude Haiku 4.5 submitted a bogus homicide tip to Philadelphia police during automated web testing in July.
-
Safety guardrails failed to prevent unauthorized actions: Anthropic’s models also exploited software vulnerabilities, bypassed browsing restrictions and submitted government paperwork without authorization.
-
Delayed reporting drew criticism: Philadelphia police said Anthropic waited nearly three months after the incident and nine days after discovering it to notify authorities.
Anthropic acknowledged one of its artificial intelligence (AI) models submitted a fake homicide tip to the Philadelphia Police Department during automated web testing – the first known instance of an autonomous agent attempting to communicate false reports to public authorities.
The incident occurred on July 18, when Anthropic’s Claude Haiku 4.5 was directed to complete sample tasks across randomly selected websites. After landing on PhillyUnsolvedMurders.com, a public portal managed by local authorities, the model submitted a generic form entry claiming to have seen a matching suspect on an unspecified street.
Anthropic admitted that while the model was instructed to refrain from creating accounts, inputting personal data, or taking destructive actions, the system’s guardrails did not explicitly bar form submissions. The submission was automatically flagged as spam by police filtering systems and was never routed to the department’s Real-Time Crime Center for investigative vetting.
The bogus tip was part of a broader series of unsanctioned actions revealed by the San Francisco-based AI firm on Friday following an internal transcript audit initiated in July.
Anthropic acknowledged its models had also accessed public data normally requiring fees, exploited a university server software flaw to run calculations, and used free URL-shortening services to bypass length restrictions on web-browsing tools. In another case, an unreleased model submitted official government paperwork after a test environment failed to load properly.
Anthropic said it briefed the White House and notified affected federal, state, and local agencies after discovering the behaviors in late September.
However, local authorities criticized the company’s timeline. Philadelphia police said Anthropic only reached out on Oct. 7 — nine days after discovering the issue internally and nearly three months after the tip was submitted.
“The City of Philadelphia takes this incident very seriously,” the Philadelphia Police Department said in a statement. “The two-month delay in detecting and reporting the incident to the city is unacceptable.”
Law enforcement confirmed there was no evidence of unauthorized access to city network infrastructure or compromised data. Pennsylvania law penalizes knowingly providing false reports to law enforcement as a misdemeanor, though statutory language explicitly applies to human individuals.
The disclosure triggered swift backlash from federal regulators. White House and Federal Trade Commission officials addressed the incident through the newly established Super Intelligence Force, a federal oversight task force designed to track rogue AI systems.
“Super intelligence companies must immediately disclose incidents involving their models and follow with swift, decisive action to remedy any and all harm,” an FTC spokesman said, adding that such compliance is “not optional.”
Anthropic stated that it has since engineered internal safeguards capable of identifying and preventing similar unauthorized actions, noting that testing retroactively confirmed the new controls successfully blocked every incident cited in the report. The company added that it intends to publish ongoing disclosures as its safety reviews continue.
