Adversarial AI & Autonomous Agent Compromise (AML.TA0001) — Detection Coverage Assessment
Defensive Scorecard & Telemetry Metrics
Section titled “Defensive Scorecard & Telemetry Metrics”| Evaluation Dimension | Observed Telemetry / Attribute | Epistemic Status |
|---|---|---|
| Tactical Identifier | AML.TA0001 | Canonical Tactic Node |
| Tactic Designation | Adversarial AI & Autonomous Agent Compromise | Tactical Dimension |
| Referential Framework | MITRE ATLAS | Taxonomy Anchor |
| Coverage Index Score | 89/100 | Posture Quantification |
| Defensive Maturity Tier | ADVANCED | Operational Readiness |
| Monitored Techniques Ratio | 13 / 15 techniques | Coverage Completeness |
| Analytical Confidence | 97% | Multi-Sensor Attestation |
| Primary Evidence Sources | SRC-CISCO-TALOS, SRC-GITHUB-ADVISORIES, SRC-MANDIANT-GOOGLE | Multi-Source Grounding |
Monitored Techniques & Primary Detection Engines
Section titled “Monitored Techniques & Primary Detection Engines”The following core techniques within this tactic are systematically monitored by Hermes detection sensors:
| Technique ID | Technique Name | Coverage Level | Primary Detection Engine |
|---|---|---|---|
| AML.T0040 | Direct & Indirect Prompt Injection | HIGH | Hermes Guardrails / Sigma |
| AML.T0043 | AI Agent Tool & Memory Poisoning | HIGH | Chronicle YARA-L |
AML.T0048 | LLM Exfiltration via Markdown & Links | HIGH | WAF / Chronicle |
Architectural Telemetry Blind Spots
Section titled “Architectural Telemetry Blind Spots”Identified operational gaps where adversary tradecraft can evade current sensor placements:
- Blind Spot 1: Multi-turn semantic context drift where prompt injection occurs gradually over hours.
- Blind Spot 2: Zero-click inter-agent communication spoofing on private WebSocket message brokers.
Prescriptive Detection Engineering Priorities
Section titled “Prescriptive Detection Engineering Priorities”Hermes recommends the following prioritized implementation roadmaps to close defense gaps:
- Priority Action 1: Enforce strict tool parameter JSON schema validation and dual-authorization gates for high-privilege actions.
- Priority Action 2: Implement LLM output sanitize filters forbidding unauthenticated markdown remote images.