davisAIgovDavis & Associates Advisory, LLC
Connect with us

davisAIgov · AI Harness

AI Harness Detection Coverage Map

See which agent actions ten Elastic rules can observe, where the AI Harness needs its own telemetry, and what to verify before enabling automated containment.

Start with the coverage summary, select a node in the map, or review the disclaimer and validation scope.

Version 1.220 November 2026Harness v13.4 · components 6.00–6.18MITRE ATLAS 2026.10Elastic detection-rules @ c59ac34

01 · BOTTOM LINETen rules provide direct signals for three Harness components

3 / 19
Direct

Extensions and MCP (6.09), tool dispatch (6.10) and downstream systems (6.11) can produce matching process, file or network events.

5 / 19
Partial

Authorization (6.07), credential brokering (6.08), audit (6.12), oversight (6.15) and attestation (6.16) have related endpoint indicators. The control’s own decisions require separate telemetry.

11 / 19
Dark

Eleven components have no direct or proxy signal in this ten-rule subset.

These rules look for observable host activity or upstream security alerts. Elastic Defend supplies many of the signals. Some rules also accept other endpoint or Windows event sources. Consult the pinned rule definitions for each rule’s inputs and query.

An agent can use an approved tool on untrusted intent while producing ordinary-looking endpoint activity. Check application records for the content source, policy decision, delegated identity and downstream outcome.

The proposed design joins endpoint telemetry with independently collected Harness events such as policy decisions, retrieval scope, delegation and budget reservations. A versioned release manifest can anchor deterministic checks where the approved configuration and the observed runtime are comparable.

The map shows 76 distinct techniques and sub-techniques across 16 ATLAS tactics. The editorial rule crosswalk assigns 9 distinct ATLAS IDs; 7 also appear in the technique map. The pinned Elastic files publish MITRE ATT&CK tags.

How to close the telemetry gaps

Proposed implementation routes for the 11 components with no signal in this rule subset.

5 / 11
Ingest

Orchestration, model calls, guardrails, delegation and evaluation. Verify what native telemetry the deployed tools emit.

3 / 11
Insert

Retrieval, memory and budget decisions. Add an observation or enforcement point where existing telemetry is insufficient.

3 / 11
Build

Inputs, context assembly and recourse. Define and instrument the evidence needed to reconstruct those decisions.

Across all 19 components, nine are classified as ingest, six as insert and four as build. These are planning estimates based on the reference architecture. A deployed gateway, broker or audit service can change a component’s route.

Three joins drive the design: deployed runtime to approved release, delegated scope to parent authority, and untrusted content to downstream action. Validate the join keys and the loss rate before treating a missing record as a finding.

02 · THE MAPRules, Harness components and ATLAS tactics

Select a rule, component or tactic to isolate its connections. Click again or press Escape to clear. The selector provides the same detail without navigating the diagram.

AI HARNESS DETECTION COVERAGE MAP Detection rules, harness components and MITRE ATLAS tactics DavisAIgov AI Harness Reference Model v13.4 · MITRE ATLAS 2026.10 · Elastic rule sample · illustrative tooling · proposed AI-DDR HARNESS PLANE Boundary Control Data Identity & action Evidence Lifecycle SEEN BY THE TEN RULES DIRECT PARTIAL DARK T0053 technique tagged on one of the ten rules ROUTE TO COVERAGE Ingest: signal exists Insert: needs a control point Build: define a new record DETECTION RULES Ten Elastic rules selected for review AI HARNESS COMPONENTS Nineteen components in six planes, with a documented example for each MITRE ATLAS TACTICS Sixteen tactics, with the techniques that land on the harness R1 HIGH risk 73 Potential Destructive AWS CLI Command Executed by GenAI Agent T0053 T0101 R2 MEDIUM risk 47 GenAI Process Connection to Suspicious Top Level Domain T0072 R3 MEDIUM risk 47 Unusual Process Modifying GenAI Configuration File T0081 R4 CRITICAL risk 99 Elastic Defend Alert from GenAI Utility or Descendant T0053 T0011 R5 HIGH risk 73 GenAI Process Accessing Sensitive Files T0085 T0085.001 T0055 R6 MEDIUM risk 47 GenAI Process Compiling or Generating Executables T0053 R7 MEDIUM risk 47 GenAI CLI Started with Unsafe Permission Bypass T0081 T0053 R8 LOW risk 21 · building block GenAI or MCP Server Child Process Execution T0053 R9 MEDIUM risk 47 GenAI Process Performing Encoding/Chunking Prior to Network Activity T0086 R10 MEDIUM risk 47 GenAI Process Connection to Unusual Domain T0072 BOUNDARY untrusted 6.01 Inputs & external content DARK Microsoft Azure Prompt Shields BUILD CONTROL PLANE decides 6.02 Orchestrator / agent loop DARK OpenTelemetry agent span conventions INGEST 6.05 Context assembly DARK Langfuse Prompt Management BUILD 6.06 Input / output guardrails DARK Amazon Bedrock Guardrails INGEST 6.00 Generative model DARK OpenTelemetry model inference conventions INGEST DATA PLANE carries 6.03 RAG subsystem DARK Elastic document-level security INSERT 6.04 Memory & state DARK LangGraph persistence INSERT 6.11 Sandbox / downstream systems DIRECT AWS CloudTrail INGEST IDENTITY & ACTION PLANE enforces 6.07 Policy / authorization gate PARTIAL Open Policy Agent decision logs INSERT 6.08 Credential broker PARTIAL HashiCorp Vault audit devices INGEST 6.09 Extensions / MCP / skills DIRECT Docker MCP Gateway INSERT 6.10 Tool dispatcher DIRECT OpenTelemetry GenAI tool execution INGEST EVIDENCE PLANE proves 6.12 Observability & audit PARTIAL Elastic Distribution of OpenTelemetry Collector INGEST LIFECYCLE PLANE before and after runtime 6.13 Identity & delegation DARK IETF RFC 8693 OAuth 2.0 Token Exchange INGEST 6.14 Evaluation & conformance DARK Promptfoo evaluation docs INGEST 6.15 Human oversight console PARTIAL LangChain human-in-the-loop INSERT 6.16 AI BOM & attestation PARTIAL CycloneDX ML-BOM BUILD 6.17 Authority & consumption DARK Cloudflare AI Gateway spend limits INSERT 6.18 Transparency & recourse DARK NIST Generative AI Profile BUILD Reconnaissance AML.TA0002 AML.T0006.003 Active Scanning: Probe AI Agent Trigger ChannelsT0006.003 AML.T0064 Gather RAG-Indexed TargetsT0064 Resource Development AML.TA0003 AML.T0002.002 Acquire Public AI Artifacts: AI Agent ConfigurationT0002.002 AML.T0060 Publish Hallucinated EntitiesT0060 AML.T0115.002 Publish Poisoned AI Artifacts: AI Agent ToolsT0115.002 Initial Access AML.TA0004 AML.T0010.003 Supply Chain Compromise: Model RepositoryT0010.003 AML.T0010.005 Supply Chain Compromise: AI Agent ToolT0010.005 AML.T0078 Drive-by CompromiseT0078 AML.T0093 Prompt Infiltration via Public-Facing ApplicationT0093 AML.T0132 Misconfigured or Publicly Exposed AI ServicesT0132 AI Model Access AML.TA0000 AML.T0040 AI Model Inference API AccessT0040 Execution AML.TA0005 AML.T0011.002 User Execution: Poisoned AI Agent ToolT0011.002 AML.T0050 Command and Scripting InterpreterT0050 AML.T0051.001 LLM Prompt Injection: IndirectT0051.001 AML.T0051.002 LLM Prompt Injection: TriggeredT0051.002 AML.T0053 AI Agent Tool InvocationT0053 AML.T0100 AI Agent ClickbaitT0100 AML.T0103 Deploy AI AgentT0103 Persistence AML.TA0006 AML.T0018.003 Manipulate AI Model: Modify Prompt Construction LogicT0018.003 AML.T0020 Training Data PoisoningT0020 AML.T0061 LLM Prompt Self-ReplicationT0061 AML.T0070 RAG PoisoningT0070 AML.T0080.000 AI Agent Context Poisoning: MemoryT0080.000 AML.T0080.001 AI Agent Context Poisoning: ThreadT0080.001 AML.T0081 Modify AI Agent ConfigurationT0081 AML.T0110 AI Agent Tool PoisoningT0110 AML.T0142 Event Triggered ExecutionT0142 AML.T0145 Poison AI Agent Learning EnvironmentT0145 Privilege Escalation AML.TA0012 AML.T0012 Valid AccountsT0012 AML.T0053 AI Agent Tool InvocationT0053 AML.T0054 LLM JailbreakT0054 AML.T0105 Escape to HostT0105 Defense Evasion AML.TA0007 AML.T0054 LLM JailbreakT0054 AML.T0067 LLM Trusted Output Components ManipulationT0067 AML.T0067.000 LLM Trusted Output Components Manipulation: CitationsT0067.000 AML.T0068 LLM Prompt ObfuscationT0068 AML.T0071 False RAG Entry InjectionT0071 AML.T0073 ImpersonationT0073 AML.T0074 MasqueradingT0074 AML.T0081 Modify AI Agent ConfigurationT0081 AML.T0092 Manipulate User LLM Chat HistoryT0092 AML.T0094 Delay Execution of LLM InstructionsT0094 AML.T0107 Exploitation for Defense EvasionT0107 AML.T0109 AI Supply Chain Rug PullT0109 AML.T0111 AI Supply Chain Reputation InflationT0111 AML.T0129 Triggers in Multimodal InputsT0129 AML.T0134 AI Targeted CloakingT0134 AML.T0138 Indicator RemovalT0138 Credential Access AML.TA0013 AML.T0055 Unsecured CredentialsT0055 AML.T0082 RAG Credential HarvestingT0082 AML.T0083 Credentials from AI Agent ConfigurationT0083 AML.T0098 AI Agent Tool Credential HarvestingT0098 Discovery AML.TA0008 AML.T0069.000 Discover LLM System Information: Special Character SetsT0069.000 AML.T0084.001 Discover AI Agent Configuration: Tool DefinitionsT0084.001 AML.T0133 Discover AI Agent Runtime CapabilitiesT0133 Lateral Movement AML.TA0015 AML.T0091.000 Use Alternate Authentication Material: Application Access TokenT0091.000 AML.T0122 Exploitation of Remote ServicesT0122 AML.T0143 LLM Prompt PropagationT0143 Collection AML.TA0009 AML.T0085.000 Data from AI Services: RAG DatabasesT0085.000 AML.T0085.001 Data from AI Services: AI Agent ToolsT0085.001 AML.T0085.002 Data from AI Services: Conversation Context and MemoryT0085.002 AML.T0085.003 Data from AI Services: System InstructionsT0085.003 AML.T0137.003 Training Data Information from AI Models: Extract Memorized DataT0137.003 AI Attack Adaptation AML.TA0001 AML.T0065 LLM Prompt CraftingT0065 AML.T0118.001 Autonomous AI Agent Communication: Direct Agent CommunicationT0118.001 AML.T0124 Autonomous Attack OrchestrationT0124 Command and Control AML.TA0014 AML.T0072 Cyber Communication ChannelT0072 Exfiltration AML.TA0010 AML.T0025 Exfiltration via Cyber MeansT0025 AML.T0077 LLM Response RenderingT0077 AML.T0086 Exfiltration via AI Agent Tool InvocationT0086 AML.T0136 Exfiltration via AI User InterfaceT0136 Impact AML.TA0011 AML.T0029 Denial of AI ServiceT0029 AML.T0031 Erode AI Model IntegrityT0031 AML.T0034 Cost HarvestingT0034 AML.T0034.002 Cost Harvesting: Agentic Resource ConsumptionT0034.002 AML.T0048 External HarmsT0048 AML.T0048.000 External Harms: Financial HarmT0048.000 AML.T0101 Data Destruction via AI Agent Tool InvocationT0101 AML.T0130 AI Agent Response BiasingT0130 v1.2 · 20 November 2026 · Harness v13.4 (6.00–6.18) · ATLAS 2026.10 · Elastic detection-rules @ c59ac34 davisAIgov · AI Harness

Nothing selected. All 19 components, ten rules and sixteen tactics are shown.

The left column shows the ten rules. The center groups Harness components by function and shows their proposed telemetry routes. The right column lists the mapped ATLAS tactics and techniques. Violet chips mark davisAIgov’s editorial mappings for the ten rules. The rule table and component table provide the same relationships in text.

03 · THE TEN RULESWhat each rule sees, and where it stops

Rule name, severity, default risk score and query type come from the pinned Elastic rule files. The published tags are MITRE ATT&CK. ATLAS links and Harness coverage grades are davisAIgov’s interpretation.

#Rule and what it seesSeverityQuery type · data scopeATLAS crosswalk · davisAIgovComponentsWhere it stops
R1Potential Destructive AWS CLI Command Executed by GenAI Agent
Matches selected destructive AWS CLI commands launched from eligible process lineages.
high
risk 73
KQL query
1 listed data patterns
T0053 T01016.10 6.11 6.08Matches command patterns. Check downstream audit records to determine whether the API call succeeded and what changed.
R2GenAI Process Connection to Suspicious Top Level Domain
Flags eligible process connections to a listed top-level domain.
medium
risk 47
EQL
3 listed data patterns
T00726.11 6.09Evaluates the configured domain suffixes. A match indicates network activity for triage and does not establish command and control.
R3Unusual Process Modifying GenAI Configuration File
Flags unusual writers of monitored GenAI configuration files.
medium
risk 47
New terms
1 listed data patterns
T00816.09 6.16 6.07Flags matching file changes. Compare the changed entry with approved configuration.
R4Elastic Defend Alert from GenAI Utility or Descendant
Surfaces qualifying Elastic Defend alerts linked to selected GenAI utilities or their process lineage.
critical
risk 99
ES|QL
Defend alerts
T0053 T00116.10 6.11 6.12Requires a qualifying underlying Elastic Defend alert. It inherits the input alert and process-lineage limits.
R5GenAI Process Accessing Sensitive Files
Flags matching access to sensitive files or creation and write events in monitored paths, subject to platform support.
high
risk 73
EQL
1 listed data patterns
T0085 T0085.001 T00556.08 6.10Observes matching endpoint file activity. Elastic documents Linux support for creation events, with access events unavailable at this revision.
R6GenAI Process Compiling or Generating Executables
Flags matching compiler or packaging activity started by selected GenAI processes.
medium
risk 47
EQL
5 listed data patterns
T00536.10 6.11Flags matching compilation or packaging activity. Assess the workflow and executable provenance.
R7GenAI CLI Started with Unsafe Permission Bypass
Flags matching command-line permission bypass or automatic approval flags.
medium
risk 47
EQL
1 listed data patterns
T0081 T00536.07 6.15Observes matching launch arguments. Verify runtime approval policy separately.
R8GenAI or MCP Server Child Process Execution
Records matching child process execution from selected GenAI or MCP activity for correlation.
low
risk 21 · building block
EQL
4 listed data patterns
T00536.10 6.09 6.12A building-block rule for matching child process events. Correlate it with task and configuration evidence.
R9GenAI Process Performing Encoding/Chunking Prior to Network Activity
Correlates eligible encoding or archive process activity with an outbound connection.
medium
risk 47
EQL
6 listed data patterns
T00866.10 6.11Correlates matching local process activity and network activity. The sequence alone cannot establish what data was sent.
R10GenAI Process Connection to Unusual Domain
Flags eligible macOS network events where a domain is new under the configured query and history.
medium
risk 47
New terms
1 listed data patterns
T00726.11 6.09 6.16Flags new terms for the configured macOS events. Review the lookback, process identity, destination and business context.

04 · COVERAGE BY COMPONENTThree tiers of detection for each of the nineteen components

Tier A describes the ten-rule sample. Tier B lists additional Elastic candidates only where a matching rule can be linked at the pinned revision. Tier C contains 39 proposed specifications. Each specification needs verified telemetry, tests and an assigned response owner before use.

ComponentA · the ten rulesB · Other Elastic candidatesC · HoundBytes AI-DDR, proposedProposed response · approval required
Boundary
6.01
Inputs & external content
Dark
Content provenance is outside the ten selected rule queries.
  • External content precedes a consequential tool call. Content tagged external enters context, and a write, send or delete follows in the same trace with no gate decision between them. Harness span + provenance tag · EQL sequence by trace T0051.001 T0093
  • Ingress anomaly on agent-read fields. Free-text fields near maximum length, cloaked text or instruction patterns in records an agent will later read. Application log, ingest pipeline · ES|QL + pattern match T0051.002 T0134
Quarantine the source record, tag it untrusted and hold dependent actions.
Control plane
6.02
Orchestrator / agent loop
Dark
Planning and agent recruitment need runtime or orchestration records.
  • Privilege-inverted agent invocation. A higher-privilege agent is invoked by a lower-privilege agent instead of by a user. Harness span (invoke_agent) · EQL T0118.001 T0143
  • Call-chain depth, fan-out or loop count above envelope. Agent-to-agent depth or iteration count exceeds the workflow's attested envelope. Harness span · ML anomaly + hard threshold T0124 T0034.002
Suspend the orchestration and withdraw the agent's discoverability.
6.05
Context assembly
Dark
Context segments and truncation require application instrumentation.
  • Protected instructions truncated or undelimited. A context-budget truncation drops system instructions, or external content sits beside them with no delimiter flag. Harness span (context lineage) · ES|QL T0068 T0069.000
  • Prompt template digest drift. The prompt construction logic in use differs from the attested template. Registry + harness span · New terms T0018.003
  • Same payload in several agents' contexts. An identical instruction string appears across agents or stores within a short window. Harness span (content hash) · ES|QL T0061 T0143
Roll the template back to its attested digest.
6.06
Input / output guardrails
Dark
Guardrail verdicts require application or provider telemetry.
  • Action on a guardrail allow with no gate decision. A consequential action follows a content-filter pass with no independent authorization record. Guardrail log + gate log · EQL sequence with missing event T0054
  • Output URL carrying data read earlier in the trace. Rendered output references an external host with encoded parameters that match retrieved record values. Harness span (output) + data lineage · ES|QL T0077 T0136
Restore the guardrail configuration and strip external references from output.
6.00
Generative model
Dark
Model identity, response and usage need inference or gateway logs.
  • Resolved model outside the approved set. The model that answered differs from the model reference in the attested release. Harness span (response model) + registry · ES|QL lookup T0010.003 T0031
  • Refusal or filter-block burst by one principal. Repeated refusals or blocked responses in one session, the usual signature of jailbreak probing. Inference log · Threshold T0054
Pin to the approved deployment and hold consequential actions.
Data plane
6.03
RAG subsystem
Dark
Verify retrieval scope and index changes with storage or application records.
  • Retrieval outside the caller's tenant or ACL. Returned object identifiers fall outside the requesting principal's namespace. Retriever log · ES|QL T0085.000 T0082
  • Unregistered or out-of-window index write. The corpus or index changes from a source that is not in the retrieval manifest, or outside a change window. Vector store audit + registry · New terms T0070 T0071
Quarantine the entries and revert to the last attested index snapshot.
6.04
Memory & state
Dark
Memory writes and checkpoints need application or store telemetry.
  • Memory write sourced from external content. A persistent memory write whose lineage traces to external-provenance content. Memory store audit + harness span · EQL sequence by trace T0080.000
  • Cross-principal memory read or bulk history deletion. A memory read across user boundaries, or mass deletion of conversation state. Memory store audit · ES|QL threshold T0085.002 T0092
Freeze memory writes and purge entries by lineage.
6.11
Sandbox / downstream systems
Direct R1 R2 R4 R6 R9 R10
Matching egress, compilation and process events can help identify downstream activity. Confirm the outcome at the target.
  • Production change from a non-production identity. A mutative statement reaches a production target from a session tagged development, or during a freeze window. Database or cloud audit + session tag · ES|QL T0101 T0048
  • Egress outside the attested allowlist. The agent or its sandbox reaches a host that is not on the registered allowlist, or one whose registration has lapsed. Elastic Defend network + registry + domain monitor · ES|QL lookup (extends R2, R10) T0072 T0025
Isolate the host or block egress, then restore from a verified backup.
Identity & action plane
6.07
Policy / authorization gate
Partial R3 R7
Configuration changes and bypass flags are proxy signals. Collect gate decisions independently.
  • Material action with no gate decision. A write, delete, send, publish or cross-privilege recruitment appears with no matching gate record in the trace. Dispatcher log + gate log · ES|QL anti-join T0053
  • Gate rule change or self-approval. The gate's policy version changes outside a change window, or an approval is issued by the orchestrator's own identity. Gate log · New terms + ES|QL (extends R7) T0081
Set the gate to default deny for mutative tools.
6.08
Credential broker
Partial R1 R5
Sensitive file activity is a proxy. Collect broker records for issued credentials and their scopes.
  • Credential used beyond task scope. One token touches more targets than the task declared, or is used outside its runtime. Broker log + target audit logs · ES|QL T0091.000 T0098
  • Secret read followed by outbound request. An agent process reads a credential store and then calls a metadata service or an external host. Elastic Defend · EQL sequence (extends R5) T0055 T0083
Revoke or narrow the token and rotate any exposed secret.
6.09
Extensions / MCP / skills
Direct R2 R3 R8 R10
Matching child process, network and file events can support endpoint visibility. Registry checks need separate evidence.
  • Tool definition changed between sessions. A tool's definition or instructions differ from the last attested digest. MCP client log + registry · New terms T0109 T0110.000
  • Unregistered MCP server or instruction-bearing tool response. An MCP child process or endpoint has no registry entry, or a tool response carries instruction patterns. Elastic Defend + harness span · ES|QL lookup (extends R8) T0010.005 T0110.002
Disable the extension and pin to the attested version.
6.10
Tool dispatcher
Direct R1 R4 R5 R6 R8 R9
Matching host process activity can support dispatch visibility. Hosted tool calls need telemetry from the execution path.
  • Tool arguments carrying external content or out-of-scope targets. A write or destructive call whose arguments include external-provenance content, or a target outside the declared scope. Dispatcher log + data lineage · ES|QL T0053 T0101
  • Collect, stage and send in one lineage. A sensitive read, then encoding or archiving, then an outbound connection in one process tree or trace. Elastic Defend + harness span · EQL sequence (joins R5, R9, R10) T0085.001 T0086
Kill the process tree and block the tool class.
Evidence plane
6.12
Observability & audit
Partial R4 R8
The rules supply correlatable alerts. Assess record integrity and producer health at the collection layer.
  • Agent claim with no corroborating producer event. The agent reports tests passed or rollback complete, and no CI or dispatcher event agrees. Harness span (output) + CI and dispatcher events · ES|QL anti-join T0067
  • Producer silence or record tampering. An expected event source stops, logging is disabled, or traces and history are deleted. Collector health + platform audit · Threshold on absence T0138 T0092
Fail closed: no evidence, no consequential action. Switch to the secondary collector.
Lifecycle plane
6.13
Identity & delegation
Dark
Delegation needs identity, scope and lineage evidence.
  • Scope expansion at a delegation hop. A child agent's scope is not a subset of its parent's, or an action has no chain back to a human owner. Delegation service + harness span · ES|QL T0118.001 T0091.000
  • Child running after revocation. A child agent, token or scheduled job is still active after its parent or owner was revoked. Broker log + harness span · EQL sequence T0121 T0103
Revoke the delegation chain, including running children.
6.14
Evaluation & conformance
Dark
Conformance needs release and evaluation evidence.
  • Evaluation evidence drift. The evaluation suite, dataset or evaluator version differs from the one bound to the release, or a failed run was not routed to a release decision. CI and evaluation pipeline + registry · ES|QL T0031 T0145
  • Conformance score below floor. Grounding, accuracy or policy-conformance canaries fall under the approved threshold for consecutive windows. Evaluation pipeline · Threshold T0031 T0020
Withdraw the release approval and revert to the last conformant configuration.
6.15
Human oversight console
Partial R7
Launch arguments provide a proxy for approval settings. Collect actual reviewer decisions and interrupt outcomes.
Verify platform and available rule inventory during onboarding.
  • Approval faster than the evidence could be read. Approvals within seconds, an approver identity equal to the agent's, or an approval reused across actions or past expiry. Oversight console log · ES|QL T0067 T0073
  • Interrupt with no observed stop. A halt was issued, and execution or credential use continues past the response deadline. Console log + dispatcher + broker · EQL sequence with missing event T0121
Escalate to a second independent approver and interrupt out of band.
6.16
AI BOM & attestation
Partial R3 R10
Configuration and destination anomalies can trigger review. Compare actual runtime state with the registered baseline.
  • Runtime manifest differs from the registry. The digest of the running agent build, tool set or policy differs from the attested release. Registry + independent verifier · ES|QL lookup T0081 T0109
  • Capability with no registry entry. A tool, MCP endpoint, model or allowlist entry appears at runtime that was never registered. Elastic Defend + registry · New terms (extends R3, R10) T0010.005 T0110
Deny activation and re-attest.
6.17
Authority & consumption
Dark
Budget reservations, enforcement and reconciliation require task or ledger evidence.
  • Budget reset or dispatch without reservation. A delegation hop resets the task budget, or a tool call is dispatched with no ledger reservation. Budget ledger + dispatcher log · ES|QL anti-join T0034.002
  • Per-action ceiling exceeded. Rows changed, spend, tokens or calls for one action exceed the approved limit, or a spend limit is removed. Ledger + cloud billing and audit · Threshold T0034 T0101
Freeze the budget and cap the rate.
6.18
Transparency & recourse
Dark
Use business-process records for notices, decisions and recourse.
Verify platform and available rule inventory during onboarding.
  • Decision without a verified record. A consequential AI-assisted decision is issued with no decision record, or with one that has no verified outcome behind it. Business system + decision record store · ES|QL anti-join T0130
  • Trace content in an outbound explanation. A notice or explanation contains unredacted trace content, secrets or third-party data. DLP on decision records · Pattern match T0085.002 T0136
Withhold publication and issue a corrected notice.

05 · MARKET TOOLING BY COMPONENTReference tooling and evidence to collectAdded in v1.1

Each row pairs a Harness function with one documented example, a signal to validate and a proposed route to coverage. Actual telemetry depends on the deployed version, plan and configuration.

Component and functionDocumented exampleEvidence to validateProposed routeProposed implementation
Boundary
6.01 Dark
Inputs & external content
Brings prompts, files, web pages, email, API and tool responses, and other agents' messages into the system.
Microsoft Azure Prompt ShieldsRecord the request identifier and scanner verdict where the integration exposes them.Build
Define the event or evidence record.
  • Ingest scanner verdicts where a platform emits them.
  • Define a provenance attribute (source, trust level, ingress channel) and require client harnesses to stamp it at intake and carry it on every span.
Control plane
6.02 Dark
Orchestrator / agent loop
Plans, selects tools, and discovers and recruits other agents.
OpenTelemetry agent span conventionsCollect agent spans from instrumented runtimes. Confirm emitted fields and version per SDK.Ingest
Start with an available signal.
  • Receive spans over OTLP into Elastic.
  • Stamp each agent's privilege tier from the registry at ingest.
  • Baseline call-chain depth and fan-out per workflow, and alert on privilege-inverted invocation.
6.05 Dark
Context assembly
Joins instructions, user intent, retrieved passages, memory and tool results into one token stream.
Langfuse Prompt ManagementRecord prompt versions. Carry segment-level provenance and truncation as application fields where needed.Build
Define the event or evidence record.
  • Add context-lineage fields to the span profile: segment source, trust tag, delimiter flag, truncation event.
  • Hash each prompt template into the AI BOM so drift is a digest mismatch.
  • Match content hashes across agents to catch a propagating payload.
6.06 Dark
Input / output guardrails
Filters content going in and coming out. Advisory to the orchestrator.
Amazon Bedrock GuardrailsCapture available verdicts at the integration and correlate them to the action trace.Ingest
Start with an available signal.
  • Enable the platform integration and its prebuilt rules.
  • Correlate each guardrail allow with a gate decision, and alert when the gate record is missing.
  • Check rendered output for external URLs that carry values read earlier in the trace.
6.00 Dark
Generative model
Inference. Proposes actions and holds no authority.
OpenTelemetry model inference conventionsCapture available request and response model identifiers and token counts.Ingest
Start with an available signal.
  • Turn on the provider integration for each platform in use.
  • Join the response model to the approved-model list in the registry.
  • Alert on resolved-model drift and on refusal bursts.
Data plane
6.03 Dark
RAG subsystem
Retrieves permitted knowledge from indexed sources.
Elastic document-level securityVerify retrieval permissions and record returned object IDs with the initiating principal where policy permits.Insert
Add or validate an observation point.
  • Place a retrieval wrapper that logs principal, filter applied, returned object IDs and source.
  • Ingest index and knowledge-base change audit.
  • Keep attested index snapshots so quarantine and rollback are one action.
6.04 Dark
Memory & state
Holds thread, persistent and shared state.
LangGraph persistenceConfirm checkpoint scope and retention, then instrument reads and writes that need independent review.Insert
Add or validate an observation point.
  • Wrap memory read and write calls so each emits an event with trace ID, principal and source provenance.
  • Turn on database audit for the checkpointer store.
  • Build the purge-by-lineage playbook used in recovery.
6.11 Direct
Sandbox / downstream systems
Where code runs and where changes land.
AWS CloudTrailCheck which downstream events are logged for the enabled AWS resources and verify the action outcome.Ingest
Start with an available signal.
  • Extend from the endpoint to the effect by ingesting cloud and SaaS audit logs for each downstream target.
  • Tag sessions as development or production at the broker.
  • Load the egress allowlist into the registry and monitor each domain's registration expiry.
Identity & action plane
6.07 Partial
Policy / authorization gate
Decides independently whether a proposed action may proceed.
Open Policy Agent decision logsIngest policy decisions from participating enforcement points and redact sensitive decision inputs.Insert
Add or validate an observation point.
  • Adopt a reference architecture in which every tool call crosses a gateway with an external policy engine.
  • Ingest each decision: permit or deny, policy version, principal, tool and an argument hash.
  • Respond by pushing a forbid policy, which takes effect on the next call.
6.08 Partial
Credential broker
Issues scoped, expiring, per-call credentials.
HashiCorp Vault audit devicesCollect configured broker audit records. Confirm which lease or token identifiers can be joined to the task.Ingest
Start with an available signal.
  • Ingest broker audit logs.
  • Map each token to the task scope in its delegation envelope and alert when it touches more targets.
  • Respond by revoking the lease or token, then confirm that use has stopped.
6.09 Direct
Extensions / MCP / skills
Brings tools into the harness with their own code and their own trust.
Docker MCP GatewayCompare observed tools and server configuration with the approved catalogue for the release.Insert
Add or validate an observation point.
  • Route MCP traffic through a gateway backed by a registry.
  • Sync the catalog to an Elastic lookup index and diff tool definitions between sessions.
  • Join R8 child-process events to the registry to flag an unregistered server.
6.10 Direct
Tool dispatcher
Executes the tool call. The privileged edge.
OpenTelemetry GenAI tool executionPair tool execution spans with authorization and downstream results where equivalent identifiers can be propagated.Ingest
Start with an available signal.
  • Join execute_tool spans to endpoint process events by host, time and trace context.
  • Carry data lineage on tool arguments so a write can be tied to what was read.
  • Respond by blocking a tool class at the gateway and killing the process tree.
Evidence plane
6.12 Partial
Observability & audit
Keeps the independent record of what happened.
Elastic Distribution of OpenTelemetry CollectorValidate trace export, sampling, retention and producer-loss reporting in the configured collector.Ingest
Start with an available signal.
  • Collect outside the agent's write authority, at a sidecar or gateway.
  • Keep a mapping layer pinned to a conventions version, because the GenAI conventions are still in development status.
  • Alert on producer silence and hash records at ingest.
Lifecycle plane
6.13 Dark
Identity & delegation
Tracks who each agent acts for and how authority is handed down.
IETF RFC 8693 OAuth 2.0 Token ExchangeValidate actor and subject claims when available, and add task scope and parent IDs where the identity system does not.Ingest
Start with an available signal.
  • Ingest identity-provider logs for agent identities.
  • Rebuild the delegation chain for each trace and compute whether each child scope is a subset of its parent's. Identity products enforce scope inside their own boundary; the check across vendors is HoundBytes work.
  • Respond by revoking the agent identity and every running child.
6.14 Dark
Evaluation & conformance
Tests the exact configuration before and after release.
Promptfoo evaluation docsRetain evaluation results with the test suite, release identifier and configured pass thresholds.Ingest
Start with an available signal.
  • Emit every evaluation run as an event: suite version, dataset hash, configuration digest, result.
  • Bind the run to the AI BOM entry it tested.
  • Reuse the red-team suites to test the detections on this map.
6.15 Partial
Human oversight console
Lets a named person approve, reject or interrupt.
LangChain human-in-the-loopRecord approvals and interrupts independently with a decision ID, expiration and scope.Insert
Add or validate an observation point.
  • Define the approval event: approver identity, action, target, policy version, expiry, time to decide.
  • Run the interrupt through the broker and the gateway so it does not depend on the agent.
  • Alert on approval speed and verify the stop after every interrupt.
6.16 Partial
AI BOM & attestation
Registers and verifies the release: model, tools, prompts and policies.
CycloneDX ML-BOMUse the inventory as input to attestation, then check runtime state against what was approved.Build
Define the event or evidence record.
  • Produce an agent-level BOM at step 0 that lists the model reference, tools, prompts, policies and allowlists, and sign it.
  • Load it into an Elastic lookup index and join it to the runtime inventory from R8, R3 and the gateway.
  • This record is the baseline that every Tier C detection compares against.
6.17 Dark
Authority & consumption
Bounds calls, tokens, time and spend as shared authority.
Cloudflare AI Gateway spend limitsWhen the chosen gateway exposes usage and limits, confirm enforcement scope and correlate to the task budget.Insert
Add or validate an observation point.
  • Put a gateway in the path and propagate a task ID so spend adds up per task across children.
  • Record a ledger reservation before dispatch.
  • Set ceilings on rows changed, spend and calls per action, and respond by freezing the key.
6.18 Dark
Transparency & recourse
Connects a decision to a notice, an explanation and a route to contest.
NIST Generative AI ProfileKeep a decision record, a minimized notice and a recourse reference where appropriate for the use case.Build
Define the event or evidence record.
  • Define the decision record: decision ID, trace ID, verified outcome, notice sent, contest route.
  • Have the case system assemble it, never the agent.
  • Scan outbound explanations for trace content and secrets before release.

06 · BUILD PLANSeven workstreams for expanding telemetry and testingAdded in v1.1

Start with the signal each deployment can provide. Define the events, join keys and policy enforcement points that are missing, then test their coverage against agreed scenarios.

WorkstreamWhat exists to build onWhat HoundBytes would have to doComponents lit
1Harness span ingest
Stage 2
OpenTelemetry GenAI conventions and a collector configured for the deployment.Publish an AI-DDR span profile pinned to one conventions version. Ship Elastic ingest pipelines and data views for it. Keep content capture off by default and store hashes. Collect outside the agent's write authority.6.00 6.02 6.10 6.12
2Platform, identity and broker log packs
Stage 1
Provider audit exports, identity activity and credential broker records where enabled.Package onboarding per platform. Normalize one agent identity across sources. Rebuild the delegation chain per trace and add the scope-subset check.6.06 6.08 6.11 6.13
3Certified control points
Stage 2
An approved MCP or AI gateway, policy decision service and other enforcement points as needed.Approve gateway and retrieval control points against defined requirements. Validate parsers and instrument retrieval or memory operations where records are missing.6.03 6.04 6.07 6.09 6.15 6.17
4Attested baseline as a service
Stage 2
An asset inventory, release manifest and component-level verification record. CycloneDX ML-BOM is one inventory option.Deliver harness step 0 as a service: inventory agents, generate and sign an agent-level BOM, load registry, envelope and budget into Elastic lookup indices, and re-attest on change.6.14 6.16
5Detection content, tested
Stage 2
The 39 proposed Tier C specifications and a controlled evaluation environment.Prioritize specifications by telemetry availability. Use controlled exercises to validate expected alerts, false positives and known blind spots.All components, through Tier C
6Response connectors and verified stop
Stage 3
Response APIs, identity and policy enforcement points, and a documented approval matrix.Build a connector for each. End every playbook with a stop-state check across children, credentials, jobs and memory. Agree a pre-authorization matrix by consequence class with each client.All components, through Tier C
7The missing schemas
Stage 2
Schemas for provenance, context lineage, approval, task identity and decision records tailored to the deployment.Define these fields as part of the implementation profile. Document which fields are local extensions and how they are versioned.6.01 6.05 6.18

Implementation principles

  • Enforce authority at the action boundary with independently controlled policy.
  • Prefer scoped credentials with short lifetimes. Verify the revocation behavior in the actual deployment.
  • Pin the telemetry schema and collect only the fields the detection and evidence plan needs.
  • Measure coverage against the approved tool and agent inventory.

Where the design fits

  • Endpoint, gateway and SaaS detections can supply different evidence about the same event.
  • The proposed AI-DDR design correlates those events against an approved operating baseline and verifies the containment outcome.
  • For partner evaluation, document who controls the data, which signals each product contributes and which response actions each team can authorize.

07 · DETECTION AND RESPONSE MODELProposed detection and response operating model

AI deviation detection and response (AI-DDR) is the operating model proposed here for a HoundBytes engagement using the Harness and an Elastic Security deployment. Define the approved scope, collect independent signals and decide ahead of time which interventions responders may take. Confirm platform integrations and delivery commitments in the statement of work.

  1. Collect endpoint and Harness telemetry

    Collect eligible host, cloud and platform signals. Instrument the gate, dispatcher, broker, retriever, memory store and delegation service where needed. Plan to correlate agent and tool events with a propagated trace or other durable join key. Verify data ownership and collection outside the agent’s write permissions.

  2. Set an approved operating baseline

    Use the approved inventory and release manifest (6.16), delegation envelope (6.13), usage limits (6.17) and conformance evidence (6.14). Compare observed runtime state with the approved scope. Use observed baselines for statistical deviations once the data is adequate.

  3. Sort every signal into three classes

    • Boundary violation. Check for a material action without the required decision record or a delegated scope outside the approved envelope. Use the agreed policy and confirm relevant telemetry.
    • Behavioral deviation. Corroborate unusual tool selection, new destinations or unexpected call-chain depth before escalating.
    • Evidence integrity issue. Check digest drift, missing producers or an uncorroborated agent claim. Route to the policy owner when evidence is insufficient.

    A policy deviation can be operational error or malicious activity. Preserve the evidence needed to determine intent and act on the immediate business risk.

  4. Correlate signals into cases

    Join endpoint process lineage, tool and agent traces, identity and delegation records, and data provenance where the evidence permits. Capture the policy boundary in question and the confidence in the conclusion.

  5. Respond on a pre-authorized ladder

    The ladder is a decision template for customer approval. Verify permissions, notice, reversal paths and stop conditions before enabling automatic actions.

    LevelTriggerActionExecuted byHuman role
    0ObserveBuilding-block or low-confidence signalAnnotate the trace, raise sampling, add the event to the case graphTriage serviceNone
    1ConstrainBehavioral deviation with corroborationGate to default deny for mutative tools; narrow token scope; withdraw discoverability; cap the budgetAuthorized gate or broker, if configuredNotified
    2InterruptBoundary violation, or deviation in a high consequence classRevoke the delegation chain including children; kill the process tree; disable the extension; isolate the hostAuthorized broker, dispatcher or endpoint response connectorApproves irreversible or business-critical action under the agreed policy.
    3RecoverStop state verifiedRestore configuration to the attested digest; purge memory and index entries by lineage; rotate secrets; rerun evaluationSystem and release ownersRelease authority signs the return to service
    4GovernEvery case at level 2 or aboveDecision record, notice and recourse assessment, baseline update, board reporting where materialCase ownerAccountable owner
  6. Verify containment and document recovery

    Verify observed stop state across the components in scope, including active child agents, credentials and scheduled jobs. Check downstream outcomes and document recovery actions. Re-run the agreed evaluation before the release owner restores service.

  7. Preserve evidence for review

    Define capture, access, retention and chain-of-custody procedures. Where appropriate, record cryptographic digests and provide repeatable verification. Agree evidence handling with counsel and the system owner.

Worked example · child-process alert

Illustrative scenario adapted from the supplied map. An editor’s extension host starts a Python interpreter on a managed workstation. Elastic R8 provides a low-severity building-block signal for eligible child-process events.

  • Observe. Check whether the process lineage matches the workstation’s approved workflow.
  • Constrain if corroborated. If a related GenAI configuration change is flagged, check recorded approval and the tool registry.
  • Consider an interrupt. If related alerts suggest access to credentials and suspicious outbound activity, reconcile with downstream evidence and the approved response policy before drawing a conclusion.

A building-block signal becomes actionable when the surrounding evidence and response threshold support it.

Measures that show the program works

  • Components with at least one live detection, out of nineteen
  • Share of agents running against an attested baseline
  • Share of material actions with a gate decision record
  • Producer coverage and recorded loss, by source
  • Time to verified containment, measured to observed stop
  • False-containment rate, by ladder level

What to report to the board

  • Which material actions have verifiable decision and stop records.
  • Which components lack a validated signal and who owns remediation.
  • Which interventions have been approved and tested, and which still require case-by-case review.
  • What evidence is available for counsel, auditors or insurers, subject to access and retention rules.

08 · ENGAGEMENT PATHShadow Mode, Pattern Capture, Agentic Cutover

A proposed three-stage engagement. Use the map to record starting coverage, validate new sources and approve the response scope.

Stage 1 · Shadow Mode

See the action edge

Start in observe-only mode. Enable eligible rules, inventory agents and tools, and document which signals and permissions are actually available.

Workstream 2 v1.1

Stage 2 · Pattern Capture

Light the dark components

Instrument the gate, dispatcher and broker first. Register the baseline and test proposed rules in alert-only mode before assigning any automated action.

Workstreams 1, 3, 4, 5 and 7 v1.1

Stage 3 · Agentic Cutover

Validate interruption and recovery

Enable the approved responses only after exercising permissions, stop-state verification, recovery and evidence handling. Reserve irreversible actions and return-to-service decisions for the designated owners.

Workstream 6 v1.1

09 · LIMITS AND VALIDATION NOTESDisclaimer and validation scope

  • Source checks. On 20 November 2026, we compared the ten rule names, severity, default risk score, query type and published framework against the pinned Elastic files. We checked the displayed ATLAS tactic and technique identifiers and titles against MITRE ATLAS 2026.10. The mappings between rules, techniques and Harness components are davisAIgov’s analytical judgments.
  • Meaning of coverage. Direct means a rule can observe matching activity associated with a component. Partial means only a related or proxy signal is shown. Dark means the analysis found no signal in the ten-rule subset. These grades do not measure detection probability or tested effectiveness. Tooling routes are implementation estimates for the reference design.
  • Implementation status. The 39 Tier C detections, telemetry joins and response ladder are specifications. They have not been implemented or tested against production data for this document. Confirm data schemas and loss rates, validate rules on representative traffic and exercise rollback before enabling containment.
  • Vendor scope. Linked product examples document relevant features. Verify version, region, licensing, configuration, data handling and response permissions for each deployment. Named vendors are examples. No sponsorship, partnership, certification or comparative ranking is represented by inclusion.
  • Worked example and evidence. The example is illustrative. The underlying alert record was not independently available for verification. Evidence handling depends on the engagement, data sources, integrity controls and applicable requirements.
  • Use. This document supports architecture and discussion. A deployment requires its own controls assessment, legal review when applicable, and written response authorization. Review primary sources.

10 · SOURCESPrimary sources and verification anchors

Each row in Reference tooling links to its primary source. External source and site links require internet access.