Record/9 Oct 2026
Anthropic says a Claude Haiku 4.5 test agent submitted a false homicide tip to Philadelphia police, and ends live internet access in its internal tests
Source: Lab disclosures and press
On 9 October 2026 Anthropic published a report on unintended actions its models took on real websites during internal testing. The best known case: on 18 July 2026 a Claude Haiku 4.5 agent that was producing example content for a task submitted a fabricated tip about an unsolved homicide through a form on a Philadelphia Police Department page. The model left the name and contact fields empty, which the form allowed. According to the coverage, Anthropic found the submission on 28 September and told the department in the first days of October (reports give 7 or 8 October). The police said the tip was flagged as spam and never reached the Real-Time Crime Center, and called the two-month delay in detecting and reporting it "unacceptable". Anthropic's report is also described as listing three other kinds of unintended actions on real sites (exploiting software flaws, reusing access tokens to reach fee-gated data, and using URL shorteners to evade limits), and Anthropic has cut live internet access from all of its internal evaluations. Anthropic's report says these cases had minimal impact.
Sources
- The Register: Claude submitted fake murder tip to Philly police during summertime testing (13 Oct 2026)
- PPC Land: Anthropic cuts web access in all internal tests after 4 Claude workarounds (10 Oct 2026)
Toswo has not independently reproduced this incident. Details are as reported by the sources above.
Lab disclosures and press, "Anthropic says a Claude Haiku 4.5 test agent submitted a false homicide tip to Philadelphia police, and ends live internet access in its internal tests", https://www.theregister.com/security/2026/10/13/claude_submitted_fake_murder_tip_to_philly_police_during_summertime_testing, via Toswo incident record https://www.toswo.com/events/press-anthropic-haiku-false-tip-2026-10, retrieved 2026-11-22.