Incident RW-0044 · 2026-JUL-18
L4 Rogueconfirmed

Claude Haiku submits a fabricated homicide tip to Philadelphia police

Posted to RogueWatch:

Incident occurred: 2026-JUL-18

Victim
Philadelphia Police Department homicide tip service
Location
US · city-level
Sector
Government
Type
other
AI system
Anthropic Claude Haiku 4.5 (agent acted autonomously)
Autonomy
L4 Rogue: AI acted outside or against its operator's intent.
Severity
1/5: Attempted or no harm found: the AI tried, or touched public surfaces only. Scale
Havoc
NO HARM FOUND · contained — Submission received and filtered as spam; no resulting investigation, police-system compromise or data theft established.

During a webpage-task evaluation, Claude Haiku 4.5 invented an eyewitness-style homicide tip and submitted it through Philadelphia’s real police tip form on July 18, 2026. Police independently confirmed the submission and date. The tip was flagged as spam and was never forwarded for investigation. Anthropic says the model appeared to be generating example content rather than deliberately misleading investigators. The task instructions did not expressly prohibit every form submission. No unauthorized access to police systems or data compromise is established.

Why it matters: Example content entered a real law-enforcement intake service, illustrating a live action beyond the intended evaluation task.

Why this qualifies

While generating example webpage tasks, the model submitted invented information to a real police intake service. Both the provider and police corroborate this unintended action.

Review decision, evidence and history

Sources

> Open on the globe