Safety failure
Realized harm
Jiushi autonomous delivery vehicle wedged a fallen e-scooter under its wheel, then accelerated back to normal speed and dragged it through Xianyang
AIR-2025-0020 AIR record · 2025-04-jiushi-delivery-vehicle-scooter-drag
A real party was affected. This incident caused actual consequences, not a demonstrated capability.
From the supporting source
Company statements said the vehicle attempted to avoid the scooter but failed to fully recognize its shape and position, causing part of it to become caught under a wheel.ctdsb.net
The record
Disclosed
2025-04-08
Occurred
2025-04-08
Vendor
Jiushi Intelligent Technology
Product
Jiushi unmanned delivery vehicle (ZTO Express route)
Risk domain
Agent behaviour › Unsafe actuation
System type
Classical autonomy (no generative model)
Who acted
Agent acted — the delivery vehicle accelerated and dragged the scooter
Attack surface
Autonomous operations
Autonomy level
Fully autonomous
Initial vector
None autonomous action
Impact type
Property damage
Confidence
High
Submitted by
EAI Internal
Crosswalk
- OWASP Agentic Top 10 (2026)
- ASI01 – Agent Goal Hijack
- MITRE ATLAS
- AML.T0048 – External Harms
- NIST AI RMF
- MEASURE 2.6; MANAGE 1.3
- EU AI Act
- Annex III §1(c)
- ISO/IEC
- TR 24028; 24029-1; 23894
- AIUC-1
- C. Safety
- Scope
- agent misuse
- Mitigation
- Require confirmation for destructive actions; Implement dry-run modes; Use layered safety controls.
- Taxonomy
- Enkrypt AI agent risk taxonomy · Black-Box Red Teaming of Agentic AI: A Taxonomy-Driven Framework for Automated Risk Discovery
Blast radius
one e-scooter; the vendor's fleet operated in more than 20 Chinese cities
Capability abused
Defense & remediation
Recorded defense
collision detection (it fired, then was overridden by the drive policy)
Guardrail kinds
Physical safety system
Guardrail outcome
Silent
Fix shipped
vehicle manually paused by the operator and suspended pending investigation
Taxonomy & references
Curator notes
The vendor's own telemetry timeline is the useful part: at 09:11:41 the wheel caught the scooter, at 09:11:44 the vehicle detected the collision and slowed — and then accelerated back to normal speed and drove on. Detection worked; the policy that consumes it did not treat 'something is attached to me' as a stop condition. Third instance of the post-collision failure pattern in this corpus after Cruise and Starship. Verified quote is the AI Incident Database summary; underlying reporting is Chinese-language. AI Incident Database incident 1438.
Supporting sources
Cite as
Enkrypt AI. "Jiushi autonomous delivery vehicle wedged a fallen e-scooter under its wheel, then accelerated back to normal speed and dragged it through Xianyang". AIR, AIR-2025-0020. https://agentincident.vercel.app/i/AIR-2025-0020/ (accessed 2026-10-19).