Skip to content

Agent Trust Governance Alignment

Level: Advanced preview extension

Simple explanation

This page maps Agent Trust capabilities to industry security frameworks and AI governance regulations. These mappings show how the toolkit's controls align with recognized standards; they do not constitute certification or formal compliance.
Audience Security architects, compliance officers, and governance teams evaluating Agent Trust for regulated environments.
Purpose Document how Agent Trust controls map to OWASP LLM Top 10, EU AI Act, and NIST AI RMF requirements.
Scope Framework alignment tables and rationale. Out of scope: threat model (see Agent Trust Profile), integration how-to.
Success criteria Reader can reference specific Agent Trust controls when responding to compliance questionnaires or risk assessments.

OWASP Top 10 for LLM Applications

OWASP LLM Risk Agent Trust Control
LLM01 - Prompt Injection Per-action tokens prevent injected prompts from invoking unauthorized tools
LLM02 - Insecure Output Handling Request binding (req_bind) ties tokens to specific requests, preventing misuse
LLM04 - Model Denial of Service Short-lived tokens + rate limiting reduce sustained abuse
LLM07 - Insufficient AI Alignment Policy engine enforces explicit rules regardless of LLM output
LLM08 - Excessive Agency Capability scoping + delegation limits + attenuation validation bound agent authority
LLM09 - Overreliance Audit receipts with obligations support human-reviewable decision trails
LLM10 - Model Theft Tool registry with signed manifests prevents tool poisoning and schema tampering

EU AI Act (Regulation 2024/1689)

Requirement Agent Trust Feature
Art. 14 - Human oversight Audit receipts with correlation IDs + approval evidence support human review
Art. 12 - Record-keeping Receipt writer produces append-only, partitioned records per decision
Art. 9 - Risk management Policy engine implements allow/deny rules with obligations and explicit constraints
Art. 15 - Robustness Short token lifetimes + replay prevention + request binding reduce attack surface
Art. 13 - Transparency Token claims (cap, ctx, del) make agent authority and provenance machine-readable

These mappings reflect current understanding of the regulation. The EU AI Act implementing rules are still being finalized; verify applicability with legal counsel for your deployment context.

NIST AI Risk Management Framework (AI RMF 1.0)

Function Agent Trust Mapping
GOVERN Policy rules are version-controlled, deterministic, and testable. Policy decision bindings (pol_bind) create an auditable link between decisions and policy versions.
MAP Capability claims make agent authorities explicit and reviewable. Tool registry entries with signed manifests provide provenance for tool schemas.
MEASURE OpenTelemetry counters and histograms (spec Section 24.1) capture minting, verification, policy evaluation, and replay detection metrics.
MANAGE Key rotation + policy updates change trust boundaries without code changes. Security modes (Demo/Pilot/Production) enforce progressive hardening.

Governance Architecture

flowchart TD
    subgraph Controls["Agent Trust Controls"]
        Policy[Policy Engine\nAllow/Deny/Obligations]
        Tokens[Capability Tokens\nScoped + Short-lived]
        Receipts[Audit Receipts\nPartitioned + Signed]
        Registry[Tool Registry\nSigned Manifests]
    end

    subgraph Frameworks["Compliance Frameworks"]
        OWASP[OWASP LLM Top 10]
        EUAI[EU AI Act]
        NIST[NIST AI RMF]
    end

    Policy --> OWASP
    Policy --> EUAI
    Policy --> NIST
    Tokens --> OWASP
    Tokens --> EUAI
    Receipts --> EUAI
    Receipts --> NIST
    Registry --> OWASP