Human Risks from AI Agents

Harms to people from agents that drive, move, persuade, watch and act.

The OECD AI Incidents Monitor recorded 324 media-reported incidents and hazards a month in 2025, up from 92 a month in 2022. Synthetic media made up 14% of those reports and cyberattacks plus fraud about 10%.

Media-reported AI incidents and hazards: 2022 92; 2025 324; Jan 2026 435. Secondary sources reporting OECD data MEDIA-REPORTED AI INCIDENTS AND HAZARDS 2022 92 2025 324 Jan 2026 435 PER MONTH, OECD AI INCIDENTS MONITOR SECONDARY SOURCES REPORTING OECD DATA
AI incidents, 2025 (AI Index)
362
OECD monthly reports, 2025
324
Synthetic media share
14%
Registry incidents
18

Risks

11 human risks, from automated driving to multi-agent misuse

Human Risks: risks covered
RiskRegistry incidentsEvidence casesControls that bound it
Autonomy and Human Safety65Behavioral envelopes, Human oversight, Monitoring
Autonomous Crime Economy35Monitoring, Identity and attestation, Access and authorization
Physical Safety44Behavioral envelopes, Human oversight, OT and ICS integration
Manipulation and Coercion15Behavioral envelopes, Human oversight, Monitoring
Surveillance24Consent and capture, Access and authorization
Impersonation04Identity and attestation, Consent and capture
Weaponization14Behavioral envelopes, Human oversight, Monitoring
Cyber-Physical Compromise04Identity and attestation, Telemetry integrity, OT and ICS integration
Agentic Misbehavior54Behavioral envelopes, Access and authorization, Human oversight
Multi-Agent Misuse04Monitoring, Identity and attestation, Behavioral envelopes
Critical Infrastructure Compromise14OT and ICS integration, Monitoring, Access and authorization

Evidence

The lead cases span deepfake fraud, driver-assist deaths and companion chatbots

Documented cases: human Risks
DateCaseCountryWhat happenedSource
Jan 2024Arup deepfake video callChina (Hong Kong)An employee made 15 transfers totaling HK$200 million ($25 million) after a call in which every other participant was a deepfakeSource
Mar 29, 2025Xiaomi SU7 driver-assist crashChinaThree students died after the car handed control back about one second before hitting a barrierSource
Sep 2025GTG-1002 AI-run espionageChinaA Chinese state-sponsored group used Claude Code for 80% to 90% of the work against about 30 targetsSource
Jan 7, 2026Character.AI and Google settlementUnited StatesFive lawsuits over teen suicides and harms linked to companion chatbots settledSource
Jan to Apr 2026Met Police live facial recognitionUnited Kingdom1.7 million faces scanned, 44 arrestsSource

Registry

18 registry incidents involve human risks

Registry incidents meet a stricter bar than the cases above: public sources name the developer, the failed control and who found it.

Registry incidents linked to this risk
IncidentDateAgentControl that failed
Uber Test Vehicle Kills Pedestrian in TempeMar 2018Autonomous test vehicleEmergency braking left to an unmonitored backup driver
Amazon Echo Sends a Private Conversation to a ContactMay 2018Voice assistantSpoken confirmation satisfied by background speech
Alexa Suggests a Dangerous Outlet Challenge to a ChildDec 2021Voice assistantNo hazard check on answers taken from the web
Tesla Staff Share Customer Car Camera Footage2019 to 2022Vehicle camera data pipelineEmployee access to customer recordings
Air Canada Chatbot Invents a Refund RuleNov 2022Customer service chatbotNo check of answers against published policy
Cruise Robotaxi Drags a PedestrianOct 2023RobotaxiPost-collision pullover with a person under the car
Packing Robot Kills a Worker in GoseongNov 2023Industrial palletizing robotRobot could move while a technician was within reach
Tesla Recalls Autopilot After NHTSA InvestigationDec 2023Level 2 driver assistanceDriver engagement controls too weak for the system's reach
Xiaomi SU7 Crashes on Assisted DrivingMar 2025Level 2 driver assistanceConstruction-zone handling with a one-second handover
Claude Opus 4 Blackmails an Engineer in a TestMay 2025Frontier model in an agent testAlignment against self-preservation
Replit Agent Deletes the SaaStr DatabaseJul 2025Coding agentNo split between development and production; destructive commands ran during a code freeze
Apollo Go Robotaxi Falls Into a Construction PitAug 2025RobotaxiPerception stopped short of a marked work zone
Stolen Drift Tokens Drain Salesforce DataAug 2025AI chat agent integrationBroad OAuth tokens held by an AI chat agent
Nx Malware Turns AI Coding Tools Against Their UsersAug 2025Coding agents (Claude Code, Gemini CLI, Amazon Q)Permission prompts skipped by command-line flags
Chinese State Group Runs Espionage Through Claude CodeSep 2025Coding agentMisuse safeguards bypassed by splitting attacks into routine-looking tasks
Waymo Robotaxis Pass Stopped School BusesSep 2025RobotaxiSchool bus stop behavior in the driving software
Cursor Agent Deletes the PocketOS DatabaseApr 2026Coding agentFull-permission API token and a delete with no confirmation
OpenAI Test Models Breach Hugging FaceJul 2026Frontier models in evaluationSandbox egress through a zero-day in a package proxy

Rules

The EU bans the worst practices outright; China regulates by service type

Rules that address human risks
RuleJurisdictionWhat it requiresStatus
EU AI Act Article 5EUBans manipulative AI, social scoring, untargeted face scraping, workplace emotion recognitionIn force Feb 2, 2025
AI-generated content labeling measuresChinaVisible and metadata labels on AI contentEffective Sep 1, 2025
Anthropomorphic AI interaction measuresChinaCompanion AI safeguards against dependenceEffective Jul 15, 2026

Status colors: green in force, blue voluntary standard or framework, amber adopted and phasing in, gray draft or pending, red withdrawn.

Record

Page Record

Risk group
Human Risks
Datasets
ai_risk_evidence.csv (5 rows); registry incidents (18 linked); ai_attack_surface.csv
Key sources
Arup deepfake video call
Xiaomi SU7 driver-assist crash
GTG-1002 AI-run espionage
Character.AI and Google settlement
Met Police live facial recognition
Last verified
October 2026