# PIR-2026-0011 - Cruise robotaxi drags a pedestrian; false crash reporting kills the business - `id`: PIR-2026-0011 - `title`: Driverless Cruise AV drags a trapped pedestrian ~20 feet in a programmed pullover; regulator penalties and shutdown follow the reporting cover-up - `date_occurred`: 2023-10-02 (San Francisco) - `date_detected`: 2023-10-02 (immediate; the dragging component surfaced publicly only via later investigation) - `date_disclosed`: incident same day; DMV permit suspension 2023-10-24; CPUC $112,500 penalty 2024; pedestrian settlement reported 2024-05; NHTSA $1.5M consent order 2024-09-30; DOJ $500K criminal resolution 2024-11-14; GM ended the robotaxi program 2024-12 - `status`: corroborated (DMV, CPUC, NHTSA, and DOJ actions all on public record) ### The agent - `agent_description`: Cruise (GM) driverless Level-4 robotaxi fleet operating commercially in San Francisco with no safety driver. - `operator_type`: enterprise - `autonomy_level`: fully-autonomous (no human in the vehicle loop) - `model_stack`: proprietary AV perception/planning stack (versions not public) - `harness`: Cruise AV platform on modified Chevrolet Bolt - `failure_locus`: agent-reasoning (perception failed to register the pedestrian beneath the vehicle; planner executed a programmed pullover on top of her) ### Authority - `authority_scope`: physical systems (multi-ton vehicle in public streets, fleet scale) - `funds_at_risk_usd`: unknown (bodily-injury liability of a driverless fleet; effectively unbounded per-incident) - `blast_radius`: public (pedestrians; and, via the fallout, the entire fleet and business) ### The failure - `root_cause`: plain-error - primary: no adversary, no malfunction of intent - wrong action from incomplete perception. Contributing: policy-violation (corporate: incomplete/false crash reporting to regulators, which drove most of the penalties) - `mechanism`: A human-driven car struck a pedestrian and threw her into the path of the driverless Cruise AV. The AV braked hard on impact, but its systems did not detect that she was trapped underneath; it then executed a programmed pullover maneuver, dragging her ~20 feet and causing severe injuries. Cruise subsequently showed regulators an incomplete video omitting the dragging; the DMV suspended its permits, CPUC fined it $112,500, NHTSA imposed a $1.5M consent order for incomplete crash reports, and Cruise admitted submitting a false record to influence a federal investigation, paying a $500K criminal fine under a DOJ deferred-prosecution agreement. GM shut the robotaxi program in Dec 2024 after investing over $10B since 2017. - `adversary_present`: no (the initiating hit-and-run driver was a contributing external event, not an adversary of the agent) - `exploitation_status`: in-wild-malfunction (real malfunction, real victim; no attacker - reclassified from in-wild-exploited, which implies an adversary) ### Impact - `severity`: catastrophic (severe injury to a member of the public; terminal for the business line) - `direct_loss_usd`: ~2.1M in fines/penalties ($1.5M NHTSA + $500K DOJ + $112.5K CPUC); pedestrian settlement reported at $8M-$12M (not officially disclosed) - `indirect_loss_usd`: GM's >$10B cumulative investment in Cruise stranded by the Dec 2024 shutdown (basis: GM/press statements; the incident and cover-up were the proximate trigger of a broader retreat) - `downtime`: fleet grounded from Oct 2023; robotaxi business never returned - `data_exposure`: none ### Detection and recovery - `detected_by`: operator + third-party (immediate on-scene; the dragging omission exposed by regulator investigation and press) - `time_to_detect`: immediate for the collision; weeks for the true mechanism to reach regulators - `time_to_recover`: never (program terminated) - `remediation`: nationwide fleet grounding, software recall, leadership exits, settlement with the victim - `structural_fix`: none within the agent that mattered - the structural outcome was regulatory (consent-order oversight) and corporate (business shutdown) - `controls_that_worked`: initial hard braking on impact bounded the first collision; NHTSA standing-general-order crash reporting functioned as the accountability channel once the omission surfaced ### Legal - `liability_holder`: Cruise/GM (operator); criminal admission via deferred-prosecution agreement for the false report, not the crash itself - `precedent_set`: first criminal resolution against an AV operator for false crash reporting; canonical case of post-incident concealment costing more than the incident - `sealed_material`: settlement amount with the pedestrian not officially disclosed ### Evidence - `telemetry_grade`: operator-logs, subsequently regulator-audited (vehicle telemetry and video were operator-held; the operator's selective disclosure of that video is itself part of the incident) - `sources`: - https://www.justice.gov/usao-ndca/pr/cruise-admits-submitting-false-report-influence-federal-investigation-and-agrees-pay - https://thehill.com/policy/technology/4909054-cruise-fined-for-failure-disclose-crash/ - https://www.washingtonpost.com/technology/2024/05/15/cruise-settlement-victim-self-driving-gm/ - https://www.claimsjournal.com/news/national/2024/05/15/323682.htm - https://www.cnbc.com/2024/12/10/gm-halts-funding-of-robotaxi-development-by-cruise.html - `independence`: DOJ primary + multiple independent outlets - strong. - `confidence`: high on penalties and mechanism (regulator orders); medium on settlement figure (reported range, never official). ### Verification notes Two corrections to the intake record. (1) The candidate stated the pedestrian settlement was "reportedly tens of millions"; contemporaneous reporting (Washington Post, Claims Journal, May 2024) put it at $8M-$12M - recorded as that range. (2) The candidate's "~$8-10B cumulative GM investment written off" understates the reported figure: GM had invested over $10B in Cruise since 2017 ("written off" is loose - the unit was absorbed into GM, not zeroed). Also made explicit that the NHTSA/DOJ/CPUC penalties punished the reporting conduct, not the dragging itself - the fine understates the underwriting fact in the opposite direction from most legal-cluster cases.