ai-safe2-framework

AISM Agent Threat Control Matrix

AISM Agent Threat and Control Matrix

Documentation updated: 2026-10-02. Clarifies the existing Sovereignty Score and introduces TCP and Evidence Assurance v1.0 as separate assessment methods. See the assessment guide for the three instruments, Technology Card workflow, and worked example. Original methodology dates below identify their baseline; this update does not change scoring formulas.

Framework: AI SAFE² v3.1 Organization: Cyber Strategy Institute Version: April 2026


Emerging evidence and technology mapping

Use the Technology Contribution Profile to record whether a source implements, validates, evidences, or challenges each mapped control. The nine-example review adds candidates for composed capabilities, forged process evidence, and learned-policy poisoning tests. These are review targets under existing canonical IDs, not new threat/control totals or completed Challenge Lab results.

Overview

Agentic AI systems introduce a threat landscape that does not exist in traditional software environments. When an AI system can autonomously call tools, spawn sub-agents, retrieve external data, and execute multi-step workflows, the attack surface expands dramatically. Threats that are theoretical for a chatbot are operational for an autonomous agent.

This document maps the primary threat categories facing agentic AI deployments to the AISM controls that address them, with cross-references to MITRE ATLAS techniques and OWASP LLM vulnerabilities. It is organized by threat category, covering both classic AI security threats and the emerging threat landscape specific to agentic systems.

v3.1 alignment: This matrix retains the AI SAFE² Cross-Pillar Governance OS (CP.1–CP.10) and adds the v3.1 protocol-governance context: CP.5.MCP profile controls MCP-1 through MCP-19, agent-to-tool enforcement, session-independent state handling, capability and audience binding, catalog/schema provenance, and extension negotiation. CP.7 Deception & Active Defense, CP.9 Agent Replication Governance, and CP.10 HEAR Doctrine remain referenced in the threat categories they directly address. T11 Multi-Turn Behavioral Conditioning covers the distinct attack class addressed by CP controls and S1.6.

This matrix is designed for security engineers, red teams, and governance practitioners who need to understand not just what controls exist, but why they exist and what specific threats they mitigate.


Threat Categories at a Glance

# Threat Category Primary Attack Surface AISM Pillars CP Governance OS MITRE ATLAS OWASP LLM
T1 Prompt Injection Input boundaries, tool responses, retrieved documents P1 Shield, P2 Ledger CP.7 Active Defense AML.T0051 LLM01, LLM02
T2 Multi-Agent Exploitation Agent-to-agent communication, orchestration trust P1 Shield, P3 Circuit Breaker, P4 Command Center CP.4 Control Plane, CP.9 ARG Agent subdomain (Oct 2025) LLM08
T3 Memory and Context Poisoning RAG pipelines, vector stores, agent memory P1 Shield, P2 Ledger CP.7 RAG canaries AML.T0000.001, AML.T0001 LLM03
T4 Supply Chain Compromise Model artifacts, dependencies, fine-tuning data, platform config P1 Shield, P2 Ledger None AML.T0002, AML.T0005, AML.T0010 LLM05, LLM10
T5 Non-Human Identity Abuse Service accounts, API keys, machine identities P1 Shield, P2 Ledger, P3 Circuit Breaker None None currently LLM06
T6 Runaway Autonomy Agent action scope, recursion, resource consumption P3 Circuit Breaker, P4 Command Center CP.8 Catastrophic Risk, CP.10 HEAR AML.T0040 LLM04, LLM08
T7 Data Exfiltration AI output channels, tool calls, agent memory P1 Shield, P2 Ledger None AML.T0057 LLM06
T8 Model Inversion and Extraction Inference API access, output analysis P1 Shield, P2 Ledger None AML.T0005 LLM10
T9 Adversarial Inputs Training data, inference inputs P1 Shield None AML.T0051, AML.T0057 LLM01, LLM03
T10 Insider and Operator Threats Governance bypasses, privilege abuse, training manipulation P2 Ledger, P4 Command Center, P5 Learning Engine CP.1 Failure Taxonomy AML.T0010 LLM09
T11 Multi-Turn Behavioral Conditioning Multi-session context, few-shot implanting, role confusion P1 Shield, P2 Ledger CP.2 Threat Model None current LLM01

T1: Prompt Injection

What it is: An adversary embeds instructions within inputs that override the agent’s intended behavior. Direct injection occurs in user inputs. Indirect injection occurs when malicious instructions are embedded in documents, web pages, tool responses, or other data the agent retrieves and processes.

Why it is especially dangerous for agents: A traditional LLM that receives a prompt injection may produce a bad output. An autonomous agent that receives a prompt injection may execute malicious tool calls, exfiltrate data, spawn additional agents, or take actions that cannot be undone.

Attack vectors:

AISM Controls:

Control Pillar Description
Adversarial prompt detection with semantic analysis P1 Shield Detect injection patterns across all input endpoints, not just user inputs
Input validation schema enforcement P1 Shield Reject inputs that do not conform to expected structure
Tool response sanitization P1 Shield Treat all tool responses as untrusted inputs; apply same sanitization as user inputs
Indirect injection surface coverage (S1.10, v3.0) P1 Shield Enumerate every non-prompt input channel — documents, emails, tool outputs — and apply identical sanitization rigor
Semantic isolation boundary enforcement (S1.3, v3.0) P1 Shield Architecturally separate trusted system instruction context from untrusted content processing at the inference layer
Agent action logging with full context P2 Ledger Record every agent action with the input that triggered it for post-incident analysis
Behavioral baseline anomaly detection P2 Ledger Alert when agent behavior deviates from established patterns after potential injection
Human approval gates for high-consequence actions P4 Command Center Require human review before irreversible agent actions regardless of trigger source
CP.7 Deception & Active Defense Layer CP Governance OS Canary documents seeded in RAG corpora detect retrieval-based indirect injection; honeypot tool endpoints identify tool squatting; first deception-class control in any AI governance framework

MITRE ATLAS: AML.T0051 (Prompt Injection) OWASP LLM: LLM01 (Prompt Injection), LLM02 (Insecure Output Handling)


T2: Multi-Agent Exploitation

What it is: Attacks that target the trust relationships, communication protocols, and coordination mechanisms between agents in a multi-agent system. When agents can instruct each other, a compromised agent becomes an attack vector against all agents it can reach.

Why it is especially dangerous for agents: Single-agent systems have a single attack surface. Multi-agent systems have an attack surface that scales with the number of agents and inter-agent communication paths. A single compromised orchestrator agent can potentially compromise all worker agents it coordinates. When orchestrators can spawn sub-agents dynamically, the attack surface expands with every spawned delegation hop.

Attack vectors:

AISM Controls:

Control Pillar Description
Agent-to-agent communication isolation in dedicated network zones P1 Shield Prevent unauthorized inter-agent communication paths
A2A protocol authentication, authorization, and encryption P1 Shield Every agent-to-agent message must be authenticated and authorized
Automated agent quarantine on behavioral anomaly detection P1 Shield Isolate agents showing anomalous behavior before they can propagate compromise
P2P agent trust scoring with reputation weighting P1 Shield Track agent behavior history and apply trust weights to agent communications
Consensus voting audit trails for multi-agent systems P2 Ledger Record how multi-agent consensus decisions are reached for audit and anomaly detection
Dynamic agent state inventory (A2.4, v3.0) P2 Ledger Real-time registry of every deployed agent with owner_of_record, ACT tier, tool authorizations, and control_plane_id
Centralized kill switch for multi-agent systems P3 Circuit Breaker Halt entire agent swarm from a single control point when needed
Multi-agent cascade containment (F3.5, v3.0) P3 Circuit Breaker Failed agents isolated at their boundary — downstream agents receive clean error signals rather than propagating failure
Consensus failure escalation to human operators P3 Circuit Breaker Surface unexpected consensus failures for human review immediately
Human approval gates for multi-agent consensus decisions P4 Command Center Require human authorization for consequential multi-agent decisions
Distributed agent health monitoring P4 Command Center Monitor individual agent health within multi-agent deployments
CP.4 Agentic Control Plane Governance CP Governance OS Single authoritative policy layer over agent identity, dynamic permission enforcement, and orchestration boundaries — board-visible governance artifact; A2A, MCP, and ACP protocol meshes evaluated against CP.3-CP.7
CP.9 Agent Replication Governance (ARG) CP Governance OS First published governance standard for agent replication: replication authority in deployment manifests enforced at gateway; ephemeral credentials with scope narrowing per delegation hop; cryptographic lineage tokens; ACT-3 max 2 hops, ACT-4 max 3 hops; 500ms full delegation-tree severance on kill signal

MITRE ATLAS: Agent subdomain techniques (updated October 2025) OWASP LLM: LLM08 (Excessive Agency)


T3: Memory and Context Poisoning

What it is: Attacks that corrupt the information an agent retrieves, stores, or uses for context. This includes RAG (Retrieval-Augmented Generation) poisoning, vector store manipulation, long-term memory corruption, and context window injection through retrieved content.

Why it is especially dangerous for agents: Agents that rely on external memory and retrieval pipelines inherit the security posture of every data source they retrieve from. A poisoned vector store or corrupted RAG corpus can cause an agent to behave incorrectly across thousands of interactions without any individual input appearing adversarial.

Attack vectors:

AISM Controls:

Control Pillar Description
Semantic similarity analysis for gradual poison injection detection P1 Shield Detect slowly introduced adversarial content through baseline comparison
RAG poisoning detection with baseline embedding monitoring P1 Shield Monitor embedding space for unexpected shifts indicating corpus corruption
Thread injection prevention with per-agent session isolation P1 Shield Isolate agent context between sessions to prevent cross-session contamination
Memory governance boundary controls (S1.5, v3.0) P1 Shield Every write to persistent agent memory requires authorization, sanitization, and an append-only audit log entry — prevents unauthorized corpus modification
SHA-256 hashing of agent state and context for integrity verification P1 Shield Detect unauthorized modifications to agent context
Periodic RAG content audits with semantic similarity checks P2 Ledger Regularly audit retrieval corpora for injected adversarial content
RAG corpus diff tracking (A2.6, v3.0) P2 Ledger Hash-verified change log for the retrieval layer; automatically correlates behavioral changes to corpus changes
AgentPoison trigger phrase detection P2 Ledger Scan retrieval corpora and outputs for known trigger phrase patterns
Baseline embedding space monitoring P2 Ledger Alert on unexpected shifts in the embedding space of retrieval stores
Memory poisoning incident response playbook P3 Circuit Breaker Define specific procedures for containing and recovering from memory poisoning events
RAG content quarantine and restoration procedures P3 Circuit Breaker Procedures for isolating corrupted retrieval content and restoring clean baselines
Context consistency verification for agent sessions P4 Command Center Monitor for context anomalies that may indicate poisoning
CP.7 Deception & Active Defense Layer CP Governance OS Canary documents seeded throughout RAG corpora detect corpus access by unauthorized agents or unusual retrieval patterns; embedding drift canaries detect gradual poisoning before behavioral change becomes detectable

MITRE ATLAS: AML.T0000.001, AML.T0001 OWASP LLM: LLM03 (Training Data Poisoning)


T4: Supply Chain Compromise

What it is: Attacks that target the pipeline from model provider to production deployment. This includes model artifact tampering, malicious fine-tuning datasets, compromised dependency packages, counterfeit model distribution, and — in v3.0 — attacks on the configuration APIs of managed cloud AI platforms.

Why it is especially dangerous for agents: An agent built on a compromised model or dependency inherits whatever backdoors or vulnerabilities were introduced upstream. These compromises are particularly difficult to detect because the agent may appear to function normally for all standard use cases while behaving maliciously in specific triggered scenarios. Cloud AI platform configuration attacks (e.g., Bedrock UpdateGuardrail API manipulation) extend the supply chain attack surface to managed services.

Attack vectors:

AISM Controls:

Control Pillar Description
Cryptographic verification of imported models and datasets (SHA-256) P1 Shield Verify model artifact integrity at every stage from download through deployment
OpenSSF Model Signing (OMS) verification at model load time P1 Shield Cryptographically verify model provenance from source through fine-tuning to deployment
Provenance chain verification from base model through fine-tuning P1 Shield Document and verify every step in the model development lifecycle
Automated SBOM scanning with CVE correlation P1 Shield Continuously scan all model dependencies against known vulnerabilities
SBOM generation and maintenance for all models and applications P2 Ledger Maintain comprehensive inventory of all model dependencies
Supply chain artifact audit and provenance documentation P2 Ledger Record complete provenance chains for all AI artifacts
Model lineage provenance ledger (A2.3, v3.0) P2 Ledger Cryptographic chain of custody from base model through every fine-tuning stage to production; extends OpenSSF OMS
SBOM version control and audit trails P2 Ledger Track all changes to software bills of materials over time
Certificate expiration monitoring for model signing P2 Ledger Alert before model signing certificates expire to prevent enforcement gaps
Automated OMS signature verification in CI/CD pipelines P2 Ledger Enforce model signing verification at deployment time, not just at load time
Cloud AI platform-specific monitoring (M4.8, v3.0) P4 Command Center Monitor Bedrock UpdateGuardrail and UpdateDataSource APIs; Azure AI Foundry configuration changes; platform attack paths that standard CloudTrail monitoring misses
Platform-specific agent security profiles (CP.5, v3.0) CP Governance OS Per-platform companion documents for Bedrock, Azure AI Foundry, n8n, LangGraph, AutoGen, and CrewAI mapping SAFE² controls to platform CVEs and monitoring telemetry

MITRE ATLAS: AML.T0002, AML.T0005 (Backdoor ML Model), AML.T0010 (Craft Adversarial Data) OWASP LLM: LLM05 (Supply Chain Vulnerabilities), LLM10 (Model Theft)


T5: Non-Human Identity Abuse

What it is: Attacks targeting the machine identities, service accounts, API keys, and automated credentials that AI systems use to access resources. As AI agents gain more autonomous capability, they require more powerful credentials, making NHI security increasingly critical.

Why it is especially dangerous for agents: Autonomous agents require credentials to call APIs, access databases, invoke tools, and communicate with other services. An attacker who obtains an agent’s credentials can act as that agent, potentially with the same broad permissions the agent requires for its legitimate operations.

Attack vectors:

AISM Controls:

Control Pillar Description
Secret scanning in all AI outputs for credentials, API keys, tokens P1 Shield Prevent agents from inadvertently exposing credentials in outputs
GitGuardian or equivalent integration for real-time secret detection P1 Shield Catch credential exposure in AI codebases and outputs immediately
Credentials stored in secret vaults with regular rotation P1 Shield Centralize credential management and enforce rotation schedules
Automated NHI discovery across cloud, on-premises, and CI/CD P1 Shield Maintain complete visibility of all machine identities
RBAC for all NHI entities P1 Shield Enforce least-privilege access for every machine identity
Just-in-time privilege elevation P1 Shield Grant elevated permissions only when needed, not persistently
Automated decommissioning of inactive NHI (more than 90 days) P1 Shield Remove credentials for inactive agents before they become attack vectors
Dedicated NHI logging channels with real-time anomaly detection P2 Ledger Monitor machine identity activity separately with tailored anomaly thresholds
Automated alerts on NHI credential misuse P2 Ledger Alert immediately on anomalous usage patterns from machine identities
Automated NHI credential rotation P3 Circuit Breaker Rotate agent credentials automatically without manual intervention
Service account disabling with certificate revocation P3 Circuit Breaker Rapidly revoke agent credentials when compromise is suspected
NHI privilege elevation requires human review P4 Command Center Ensure human oversight of any privilege escalation for machine identities

MITRE ATLAS: No current dedicated NHI techniques; monitoring for addition OWASP LLM: LLM06 (Sensitive Information Disclosure)


T6: Runaway Autonomy

What it is: Agent behavior that exceeds intended operational boundaries, consuming excessive resources, taking unauthorized actions, or executing recursive workflows without termination. This is not always the result of an attack. It can emerge from legitimate agent goals pursued without adequate constraints.

Why it is especially dangerous for agents: An agent with broad tool access and a runaway goal can cause significant harm in a short time. Resource consumption, unauthorized API calls, data modifications, and downstream service impacts can occur faster than human observers can respond without automated containment. ACT-3 and ACT-4 deployments — autonomous and orchestrator tiers — represent the highest runaway risk profile because they operate with post-hoc human review and can spawn other agents.

Attack vectors:

AISM Controls:

Control Pillar Description
Rate limiting on all API calls, model invocations, and agent actions P3 Circuit Breaker Hard limits on action frequency regardless of agent goal state
Agent recursion limit governor (F3.2, v3.0) P3 Circuit Breaker Hard cap on tool-calling depth enforced at the API gateway layer — not the system prompt; default maximum 4; fail-closed
Kill switches accessible to operators with escalation procedures P3 Circuit Breaker Multiple activation paths for emergency agent termination
Redundant kill switches: hardware and software with multi-stage shutdown P3 Circuit Breaker Defense-in-depth for containment; no single point of failure in shutdown capability
Swarm quorum abort mechanism (F3.3, v3.0) P3 Circuit Breaker Decentralized threshold-based abort: when a configurable quorum of swarm agents agree the task should stop, coordinated shutdown proceeds without a centralized kill signal
Blast radius containment via compartmentalization P3 Circuit Breaker Limit the scope of impact from any single runaway agent
Circuit breakers with graceful degradation paths P3 Circuit Breaker Reduce autonomy before full shutdown when possible
Real-time performance dashboards for all production AI systems P4 Command Center Detect resource consumption anomalies before they become incidents
Anomaly detection with automated alerting P4 Command Center Alert on unexpected action rates, tool invocation patterns, or resource usage
Token usage and cost tracking across all AI systems P4 Command Center Detect runaway behavior through economic signals
CP.8 Catastrophic Risk Threshold Controls CP Governance OS Documents behavioral indicators that trigger emergency suspension regardless of business continuity impact — unauthorized compute acquisition, communication outside approved list, weaponizable capability; required for ACT-3/ACT-4 deployment approval
CP.10 HEAR Doctrine CP Governance OS Named Human Ethical Agent of Record with cryptographic signing key and unilateral kill-switch authority for all ACT-3/ACT-4 deployments; Class-H actions (irreversible, financially material, security-control-modifying) require HEAR signature before execution; fail-closed — no automatic approval path

MITRE ATLAS: AML.T0040 (Exploit Public-Facing Application) OWASP LLM: LLM04 (Denial of Service), LLM08 (Excessive Agency)


T7: Data Exfiltration

What it is: Adversarial techniques that cause AI agents to expose sensitive data in their outputs, tool calls, or inter-agent communications. This includes both targeted exfiltration (attempting to extract specific data) and inadvertent disclosure (agents including sensitive data in outputs without adversarial intent).

AISM Controls:

Control Pillar Description
PII/PHI masking and tokenization across all pipelines P1 Shield Prevent sensitive data from appearing in agent context or outputs
DLP controls with tokenization and pseudonymization P1 Shield Data loss prevention enforcement at the Shield layer
Secret scanning in all AI outputs P1 Shield Detect credentials, API keys, and other secrets in agent outputs before they are sent
Data isolation and access boundary enforcement P1 Shield Ensure agents can only access data appropriate for their function
User interaction logging with prompt and query capture P2 Ledger Maintain audit trail of all data exposures for investigation and compliance
Semantic execution trace logging (A2.5, v3.0) P2 Ledger Captures full agent execution trace — reasoning chain, every tool call with parameters, every memory operation — written to an append-only store the agent cannot modify
Decision traceability and provenance documentation P2 Ledger Track information flow from input through agent reasoning to output

MITRE ATLAS: AML.T0057 OWASP LLM: LLM06 (Sensitive Information Disclosure)


T8: Model Inversion and Extraction

What it is: Techniques that use the inference API to reconstruct training data, extract model weights, or clone model behavior. For agents with broad tool access, model extraction can expose proprietary systems, training data, or intellectual property.

AISM Controls:

Control Pillar Description
Rate limiting and IP/role-based API restrictions P1 Shield Limit query volume from any single source to impede systematic extraction
Whitelist-based tool access with least privilege P1 Shield Minimize the surface area available for extraction attempts
API usage quota monitoring with anomaly alerts P4 Command Center Detect systematic querying patterns characteristic of extraction attacks
Behavioral baseline anomaly detection P2 Ledger Alert on query patterns that deviate from normal usage

MITRE ATLAS: AML.T0005 (Backdoor ML Model) OWASP LLM: LLM10 (Model Theft)


T9: Adversarial Inputs

What it is: Inputs crafted to exploit model-specific vulnerabilities, including homoglyph substitution, invisible character injection, encoding attacks, and other techniques that bypass content filters while achieving adversarial goals.

AISM Controls:

Control Pillar Description
Homoglyph and invisible character injection prevention P1 Shield Normalize inputs to detect visually similar but functionally different characters
Format normalization before tokenization across all pipelines P1 Shield Standardize encoding before inputs reach model tokenization
Input encoding validation: UTF-8 enforcement P1 Shield Reject inputs with unexpected encoding that may be attempting bypass
Toxicity scoring with organizational policy enforcement P1 Shield Evaluate inputs against defined content policies
Statistical anomaly detection on input data with automated profiling P1 Shield Detect statistically unusual inputs that may indicate adversarial crafting
Adversarial input fuzzing pipeline (S1.4, v3.0) P1 Shield Integrate automated adversarial testing into CI/CD — systematic attack probe battery before every model update and deployment

MITRE ATLAS: AML.T0051, AML.T0057 OWASP LLM: LLM01 (Prompt Injection), LLM03 (Training Data Poisoning)


T10: Insider and Operator Threats

What it is: Threats originating from within the organization, including operators who abuse their access to AI systems, developers who introduce vulnerabilities intentionally, and governance bypasses that undermine control effectiveness.

AISM Controls:

Control Pillar Description
Tamper-proof, centralized logging with cryptographic signing P2 Ledger Ensure audit trails cannot be modified by insiders
Dedicated NHI logging channels with real-time anomaly detection P2 Ledger Detect insider abuse of machine identities
Decision provenance documentation from input to output P2 Ledger Create accountability trail for all AI-driven decisions
Human approval workflows for critical AI actions P4 Command Center Require cross-functional review for high-impact operations
Cross-functional AI governance collaboration P4 Command Center Distribute governance authority to prevent single-person control
Red team and adversarial testing of AI systems P4 Command Center Surface insider-accessible vulnerabilities through structured testing
Comprehensive operator training with accountability framework P5 Learning Engine Build culture of AI accountability with documented expectations
Role-based training programs P5 Learning Engine Ensure all operators understand their responsibilities and the consequences of violations
CP.1 Agent Failure Mode Taxonomy CP Governance OS Every agentic incident tagged with cognitive_surface=(model|memory|both) and memory_persistence=(session|cross_session) — separates ordinary prompt failures from belief drift and memory corruption; prevents insider incidents from being misclassified as model errors

MITRE ATLAS: AML.T0010 (Craft Adversarial Data) OWASP LLM: LLM09 (Overreliance), LLM07 (Insecure Plugin Design)


T11: Multi-Turn Behavioral Conditioning (Added v3.0)

What it is: Attacks that exploit multi-session interaction to gradually modify agent behavior through repeated exposure to crafted inputs. Unlike T1 prompt injection — which attempts a single-turn override — behavioral conditioning operates across many interactions, implanting behavioral patterns that persist across sessions and bypass per-input detection.

Why it is a distinct threat category: T1 prompt injection is detectable at the input boundary on a per-message basis. Multi-turn behavioral conditioning is not. It works by changing how an agent interprets future inputs — through few-shot pattern implanting, role confusion across conversations, or systematic bias toward specific responses. Standard injection filters do not detect it because no individual message appears adversarial. Detection requires semantic analysis across multiple sessions and behavioral baseline comparison over time.

Why it is especially dangerous for agents: Agents with persistent memory and cross-session context are uniquely susceptible. The attack surface grows with every session the agent maintains. An agent that has been successfully conditioned will continue to behave anomalously even after the attacker stops interacting with it, because the behavior is now part of the agent’s learned context rather than an active injection.

Attack vectors:

AISM Controls:

Control Pillar Description
Cognitive injection sanitization (S1.6, v3.0) P1 Shield Semantic intent analysis layer that detects multi-turn behavioral conditioning, role confusion, and few-shot pattern implanting across sessions — beyond what pattern-matching filters catch
Per-agent session isolation P1 Shield Prevent cross-session contamination of agent context; conditioning in one session cannot influence behavior in another
Memory governance boundary controls (S1.5, v3.0) P1 Shield Authorization required for every write to persistent agent memory — prevents conditioning from persisting without audit trail
Behavioral baseline anomaly detection across sessions P2 Ledger Alert when agent behavior trends over multiple sessions rather than triggering only on per-input anomalies
Semantic execution trace logging (A2.5, v3.0) P2 Ledger Cross-session reasoning chain capture enables retrospective analysis of conditioning progression
RAG corpus diff tracking (A2.6, v3.0) P2 Ledger Identifies when conditioning has succeeded in altering retrieved context rather than just immediate outputs
Behavioral drift baseline and rollback (F3.4, v3.0) P3 Circuit Breaker Establishes measurable behavioral baselines; automated rollback when drift exceeds configurable thresholds — catches successful conditioning before it becomes entrenched
Context consistency verification P4 Command Center Monitor for cross-session context anomalies that may indicate active conditioning campaign
CP.2 Adversarial ML Threat Model Integration CP Governance OS Mandatory governance artifact for ACT-2+ deployments mapping every threat with temporal_profile=(immediate|delayed_days|delayed_weeks|chronic) — explicitly designed to capture time-shifted attacks like behavioral conditioning campaigns

MITRE ATLAS: No current dedicated technique; closest is AML.T0051 (Prompt Injection) — monitoring for dedicated T11-class technique addition OWASP LLM: LLM01 (Prompt Injection) — behavioral conditioning not yet explicitly categorized


Threat Coverage by AISM Pillar

Threat Category P1 Shield P2 Ledger P3 Circuit Breaker P4 Command Center P5 Learning Engine CP Governance OS
T1: Prompt Injection Primary Supporting None Supporting Continuous improvement CP.7 deception layer
T2: Multi-Agent Exploitation Primary Supporting Primary Primary Red team CP.4 control plane, CP.9 replication governance
T3: Memory Poisoning Primary Primary Primary Supporting Research integration CP.7 RAG canaries
T4: Supply Chain Compromise Primary Primary None Supporting Threat intelligence CP.5 platform profiles
T5: NHI Abuse Primary Primary Primary Primary Training None
T6: Runaway Autonomy None Supporting Primary Primary Red team CP.8 catastrophic thresholds, CP.10 HEAR
T7: Data Exfiltration Primary Primary None Supporting Awareness None
T8: Model Extraction Supporting Primary None Primary None None
T9: Adversarial Inputs Primary Supporting None None Continuous improvement None
T10: Insider Threats None Primary None Primary Primary CP.1 failure taxonomy
T11: Behavioral Conditioning Primary Primary Supporting Supporting None CP.2 threat model

Using This Matrix

For red teams: Use the threat categories and attack vectors as the basis for adversarial test scenarios. Every attack vector listed here should have a corresponding red team exercise. Coverage gaps in the AISM self-assessment will indicate which threats are least likely to be detected. T11 requires multi-session red team exercises, not single-interaction tests.

For security architects: Use the control columns to verify that specific threats are addressed in your AISM pillar implementations. A control appearing in this matrix that is not implemented in your environment is a documented gap. CP Governance OS controls are cross-pillar and should be verified separately from pillar-level controls.

For compliance teams: Use the MITRE ATLAS and OWASP LLM cross-references to map threat coverage to threat framework requirements. Note that T5 (NHI Abuse) and T11 (Behavioral Conditioning) currently have limited MITRE ATLAS coverage — CSI is tracking both for technique addition in upcoming Atlas updates.

For governance leaders: Use the threat category overview to understand the agentic threat landscape and prioritize governance investment in the areas of highest organizational exposure. CP.8 Catastrophic Risk Thresholds and CP.10 HEAR Doctrine are deployment blockers for ACT-3 and ACT-4 systems — their absence constitutes an undocumented organizational risk acceptance.



© 2026 Cyber Strategy Institute. Licensed under CC BY 4.0.