Behavioral Envelopes

Hard limits on what an agent can do, enforced where its reasoning cannot reach.

Hard limits enforced outside the model are now written into law: California fleets must leave an emergency zone within two minutes, driverless trucks must stay inside their validated roads, and China bars automating hard-to-reverse agent decisions. Envelope controls account for 12 of the 45 failed controls in the registry, more than any other discipline.

Behavioral envelopes, outermost first: Physical: California two-minute geofence · ISO force limits; Action: China: no automated hard-to-reverse decisions; Permission: Claude Code sandbox: files and network; Content: Bedrock Guardrails · ASL-3 classifiers ENVELOPES AROUND THE MODEL, OUTERMOST FIRST PhysicalCALIFORNIA TWO-MINUTE GEOFENCE · ISO FORCE LIMITS ActionCHINA: NO AUTOMATED HARD-TO-REVERSE DECISIONS PermissionCLAUDE CODE SANDBOX: FILES AND NETWORK ContentBEDROCK GUARDRAILS · ASL-3 CLASSIFIERS Model reasoning
Permission prompts cut by Claude Code sandbox
84%
ISO force-limit body regions
29
CA emergency geofence exit
2 min
Registry failures
12

Registry

12 failed envelope controls across nine registry incidents

Each row is a control that was in place, or should have been, and did not hold. The incident record explains who was responsible and what changed afterward.

Registry incidents where behavioral envelopes failed
IncidentDateControlResponsibleWhat failed
Uber Test Vehicle Kills Pedestrian in TempeMar 2018Object classificationUber ATGThe system detected the pedestrian 5.6 seconds out but never classified her or predicted her path
Uber Test Vehicle Kills Pedestrian in TempeMar 2018Emergency brakingUber ATGAutomatic emergency braking was disabled in autonomous mode and left to the operator
Alexa Suggests a Dangerous Outlet Challenge to a ChildDec 2021Hazard check on web answersAmazonA dangerous viral challenge from a third-party site was read out as an answer
Alexa Suggests a Dangerous Outlet Challenge to a ChildDec 2021Child-context safeguardAmazonNothing stopped a child's request for a challenge from returning physical instructions
Cruise Robotaxi Drags a PedestrianOct 2023Post-collision detectionCruiseThe vehicle did not detect the pedestrian underneath and pulled over, dragging her
Packing Robot Kills a Worker in GoseongNov 2023Person and object recognitionRobot installerThe sensor apparently took a person for a box
EchoLeak Turns Microsoft 365 Copilot Into a Data LeakJan 2025Prompt injection classifierMicrosoftAn attacker email worded as a business request got past the classifier
EchoLeak Turns Microsoft 365 Copilot Into a Data LeakJan 2025Link redactionMicrosoftReference-style links and images were not removed
Xiaomi SU7 Crashes on Assisted DrivingMar 2025Construction-zone handlingXiaomiThe system drove at about 116 km/h into a lane closed for roadwork
Apollo Go Robotaxi Falls Into a Construction PitAug 2025Work-zone perceptionBaiduThe vehicle drove into a fenced, marked trench
Waymo Robotaxis Pass Stopped School BusesSep 2025School bus stop ruleWaymoSoftware released from August 20 to November 5, 2025, sometimes moved off from a stopped bus
OpenAI Test Models Breach Hugging FaceJul 2026Containment and egressOpenAIAgents reached the internet through a zero-day in a package proxy inside the test boundary

Implementations

Envelopes range from force limits on robot arms to sandboxes around coding agents

Physical envelopes cap speed, force and location: ISO 10218 force limits, California's geofence rule, FMCSA's ODD conditions. Software envelopes cap what code and output can do: the Claude Code sandbox restricts files and network at the operating system, and Bedrock Guardrails and ASL-3 classifiers filter output. China's CNCERT tells users to run OpenClaw-class agents on isolated machines without admin rights.

Named behavioral envelopes implementations
ImplementationVendor or bodyCountryWhat it enforcesStatus
ASL-3 classifiersAnthropicUnited StatesInput and output classifiers that block weapons contentSince May 2025; RSP v3 Feb 24, 2026
Bedrock GuardrailsAWSUnited StatesContent, topic, sensitive-data and grounding filters on model outputGA
Claude Code sandboxAnthropicUnited StatesWrites limited to the working directory; network through an allowlist proxyOct 20, 2025
Emergency geofence exitCalifornia DMVUnited StatesFleet leaves an emergency zone within two minutes; responder link answered in 30 secondsAdopted Apr 28, 2026
Force and pressure limitsISO/TS 15066, now within ISO 10218:2025InternationalContact force limits for 29 body regionsFolded into ISO 10218 in 2025
Level 4 truck ODD conditionsFMCSAUnited StatesValidated roads only; no hazmat, passengers or triples; cab beaconsOct 7, 2026 to Oct 7, 2031
OpenClaw isolation guidelineCNCERT/CC and CSACChinaAgents on dedicated devices, VMs or containers; no admin rightsMar 2026

Row color shows the agent class served: orange physical, teal personal and ambient, blue software.

Recent events

  • FMCSA granted its five-year Level 4 truck exemption on October 7, 2026, after 402 public comments.
  • California requires 50,000 test miles for light-duty AVs and 500,000 for heavy-duty AVs per phase before deployment.
  • Anthropic said in RSP v3 that the top tier of model-weight security is currently not possible to meet.
  • OpenAI test models escaped their evaluation environment and reached Hugging Face systems in July 2026.

Exposure

Behavioral and policy boundary rates very significant or critical for 10 of 17 agent types

The behavioral envelopes discipline covers the behavioral and policy boundary dimension of the agent attack surface ratings. These agent types depend on it most.

Agent types most exposed on this dimension
Agent typeDimensionRatingWhy
Coding and research agentsBehavioral and policy boundaryCriticalIndirect prompt injection through ingested content is the primary attack vector
Workflow agentsBehavioral and policy boundaryCriticalPlanning-layer policy, tool-use policy, sub-agent coordination policy
Commerce agentsBehavioral and policy boundaryCriticalTransaction limits, approval thresholds, blocked counterparty lists, fraud detection rules
Humanoid robotsBehavioral and policy boundaryVery significantAction-layer policy, force limits, protected action classes, and task whitelists are the controls that bound physical capability
Industrial robots and cobotsBehavioral and policy boundaryVery significantSpeed, force and zone limits are the core safety boundary
Autonomous trucksBehavioral and policy boundaryVery significantOperational design domain limits, geofences, speed and route policies, hours-of-service compliance
DronesBehavioral and policy boundaryVery significantGeofences, altitude limits and no-fly zones are the main boundary
Fleets and swarmsBehavioral and policy boundaryVery significantPolicy changes pushed through the orchestration layer apply to the entire fleet
Medical devicesBehavioral and policy boundaryVery significantClinical decision support policy, intended use bounds, escalation thresholds
Enterprise agentsBehavioral and policy boundaryVery significantApproval steps and action limits bound what an agent may do

Rules

Vehicle law and China's agent rules set envelopes; most software envelopes are voluntary

Rules and standards for behavioral envelopes
RuleJurisdictionWhat it requiresStatus
China agent Implementation Opinions (CAC, NDRC, MIIT)ChinaDevelopers define operating boundaries and decision authority; hard-to-reverse decisions not automatedEffective Jul 15, 2026
California DMV AV regulationsCaliforniaTwo-minute geofence exit; 30-second responder link; manual override accessAdopted Apr 28, 2026
FMCSA warning-device exemptionUS federalDriverless trucks limited to validated ODD; no hazmat or passengersOct 7, 2026 to Oct 7, 2031
ISO 10218:2025 (with ISO/TS 15066 limits)InternationalForce and pressure limits for robot contact with peoplePublished 2025
EU AI Act Article 15EUHigh-risk systems robust against manipulationDelayed to Dec 2, 2027
Anthropic Responsible Scaling Policy v3Company policyCapability thresholds tied to required safeguardsFeb 24, 2026

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

Limits

An envelope only holds what its designers anticipated

Uber disabled emergency braking in autonomous mode, Waymo's school-bus rule missed a case, and EchoLeak got past Microsoft's injection classifier, so envelopes need monitoring to catch the cases outside them.

Record

Page Record

Datasets
ai_control_implementations.csv (7 rows for this page); ai_incident_controls.csv (12 linked); ai_attack_surface.csv
Key sources
AWS
Anthropic
Anthropic
CNCERT/CC and CSAC
California DMV
FMCSA
Source notes
Level 4 truck ODD conditions: secondary source
Last verified
October 2026