AI Security Atlas › All entries › Multi-Agent & Memory Trust
Agent Workload Identity & Signed Tasks
Each agent holds a verifiable workload identity, and every delegated task is signed and audience-scoped.
Open in the interactive atlas →Impact
Makes impersonation and task injection between agents cryptographically detectable.
Why this severity
Guardrails take the highest severity of the risks they mitigate: Unauthenticated Agent Channels (Medium), Inter-Agent Impersonation (Medium).
Method: OWASP Risk Rating Methodology
Related
Mitigates
- Unauthenticated Agent Channels Vulnerability · Medium severity
- Inter-Agent Impersonation Threat · Medium severity
Needed by
- Tool Manifest Pinning & Review Guardrail · Fixes critical risks
- Blast-Radius Isolation & Circuit Breakers Guardrail · Fixes high risks
Remediation
- Issue SPIFFE-style identities per agent workload.
- Sign task envelopes and include audience and expiry.
- Log delegation chains for forensic replay.
How to verify
- Forged identities fail mutual TLS handshakes.
- Delegation audit reconstructs the full chain for any task.
Pattern
verifyEnvelope(msg, { audience: SELF_ID, maxAgeSec: 60 });Framework mapping
OWASP Top 10 for LLM Applications 2025
- LLM06:2025 Excessive Agency
OWASP Top 10 for Agentic Applications 2026
- ASI07:2026 Insecure Inter-Agent Communication
- ASI03:2026 Identity and Privilege Abuse
MITRE ATLAS
- AML.M0032 Segmentation of AI Agent Components
- AML.M0026 Privileged AI Agent Permissions Configuration
NIST AI RMF 1.0
- MEASURE 2.7 AI system security and resilience are evaluated and documented