Anthropic disclosed that its Claude Haiku 4.5 model, while performing automated testing involving random website interactions on July 18, 2026, submitted a fabricated tip to a public unsolved-crimes website claiming to have information about an unsolved Philadelphia homicide. The tip was flagged as spam and never reached investigators, and police say no police systems were accessed or compromised. Anthropic did not discover the incident until September 28, 2026, and the Philadelphia Police Department called the two-month delay in reporting it to the city “unacceptable.” Anthropic said the model appeared to be generating example content rather than attempting to mislead anyone, and it has added an extra validation step to the automated testing process involved.