Skip to content

COMPLIANCE.md

Latest commit

 

History

History
195 lines (148 loc) · 8.29 KB

COMPLIANCE.md

File metadata and controls

195 lines (148 loc) · 8.29 KB

Compliance Mapping

How AgentMint receipt fields and runtime controls map to regulatory frameworks.

SOC 2 Trust Services Criteria

CC6.1 — Logical and Physical Access Controls

AgentMint Feature Evidence
Scoped delegation (scope in plan) Only explicitly authorised actions proceed. Receipt proves scope was checked.
Checkpoint enforcement Sensitive actions require explicit human approval before execution.
Circuit breaker Per-agent rate limits prevent runaway access patterns.
original_approver on receipt Every action traces back to the human who authorised the plan.
key_id on receipt Signing key is identified, supporting key rotation and revocation.

CC7.2 — Monitoring of System Components

AgentMint Feature Evidence
Receipt chain (previous_receipt_hash) Immutable, hash-linked audit trail of every action.
FileSink (JSONL) Real-time append-only log compatible with SIEM ingestion.
evidence_hash (SHA-512) Tamper-evident hash of action evidence at time of signing.
policy_hash on receipt Captures exact policy state at decision time.
Session trajectory Rolling window of recent agent actions within a session.

CC8.1 — Change Management

AgentMint Feature Evidence
policy_hash on receipt SHA-256 of canonical policy (scope + checkpoints + delegates_to). Any policy change produces a different hash.
Plan signature (Ed25519) Plan content is signed at approval time. Modifications invalidate the signature.
Evidence export (zip) Self-contained verification package with VERIFY.sh script.

PI1.1 — Privacy Criteria

AgentMint Feature Evidence
Shield module (PII scanning) SSN, email, phone, credit card detection before action execution.
Shield redaction PII matches are heavily redacted in threat previews.
Scope enforcement Agents cannot access data outside their approved scope.

NIST AI Risk Management Framework (AI RMF 1.0)

MAP 1.1 — Intended purpose and context of use

AgentMint Feature Mapping
Plan issuance with action description Documents the intended purpose of each agent action.
scope definition Explicitly bounds the context of permitted operations.
delegates_to list Names authorised agents, preventing unauthorised use.
Reasoning capture (reasoning_hash) Records the agent's stated rationale at decision time.

MEASURE 2.3 — AI system performance monitoring

AgentMint Feature Mapping
Receipt chain Continuous log of every action with pass/fail status.
in_policy field Binary indicator of whether each action was within approved bounds.
Circuit breaker state Quantitative measure of agent activity rate.
Session counters Per-pattern action counts within a session.
FileSink logs Machine-parseable performance data for trend analysis.

MANAGE 3.1 — Risks are responded to

AgentMint Feature Mapping
Circuit breaker (open state) Automatic blocking when agent exceeds rate thresholds.
Checkpoint enforcement Human escalation for sensitive action categories.
Shield blocking Automatic rejection when content matches threat patterns.
Session escalation Configurable deny/escalate thresholds per action pattern.
Scope intersection (delegation) Child agents receive only the intersection of parent permissions.

GOVERN 1.1 — Policies and procedures

AgentMint Feature Mapping
policy_hash Cryptographic proof of which policy was in effect.
Plan signature Non-repudiation of policy approval.
original_approver Attribution to the human who set the policy.
Evidence export Auditor-friendly package for policy review.

HIPAA (45 CFR Part 164)

164.312(a)(1) — Access Control

AgentMint Feature Evidence
Scoped delegation Technical enforcement of minimum necessary access.
Checkpoint patterns PHI access patterns can require explicit approval.
Circuit breaker Prevents bulk data access by rate-limiting agents.
delegates_to Only named agents can act under a plan.

164.312(b) — Audit Controls

AgentMint Feature Evidence
Receipt chain Hardware-independent audit trail with hash linking.
RFC 3161 timestamps Third-party timestamp authority proves when actions occurred.
FileSink JSONL log suitable for audit retention requirements.
Evidence export Self-contained verification for compliance audits.

164.312(c)(1) — Integrity

AgentMint Feature Evidence
Ed25519 signatures Receipts are cryptographically signed; tampering is detectable.
evidence_hash (SHA-512) Integrity verification of action evidence.
output_hash Integrity verification of tool output.
previous_receipt_hash Chain linking detects insertion or deletion of records.

164.312(d) — Person or Entity Authentication

AgentMint Feature Evidence
original_approver Human identity tied to plan authorisation.
key_id Cryptographic key identity for signing entity.
Agent co-signatures Agents can provide their own key-based attestation.
Plan verification verify_plan() confirms human signature is valid.

164.312(e)(1) — Transmission Security

AgentMint Feature Evidence
Shield module Scans tool inputs and outputs for PII before transmission.
Shield blocking Prevents exfiltration of detected secrets and PII.

EU AI Act — Article 12 (Record-keeping)

Article 12 requires high-risk AI systems to include logging capabilities that enable monitoring of the system's operation throughout its lifecycle.

12.1 — Automatic recording of events

AgentMint Feature Mapping
Receipt chain Every agent action is automatically recorded with unique ID, timestamp, agent, action, and outcome.
FileSink Continuous JSONL logging of all notarised events.
Session trajectory Rolling window captures operational context.

12.2 — Traceability of system operation

AgentMint Feature Mapping
plan_id to receipt_id chain Full traceability from human approval to individual actions.
parent_plan_id (delegation) Multi-agent chains are fully traceable.
previous_receipt_hash Hash linking provides tamper-evident ordering.
RFC 3161 timestamps Third-party time proof for each event.

12.3 — Level of detail appropriate to intended purpose

AgentMint Feature Mapping
evidence dict on receipt Captures action-specific context (tool args, results).
reasoning_hash Links to agent's stated rationale without storing full text in receipt.
output_hash Captures tool output integrity without duplicating content.
policy_hash Records exact policy version at decision time.

12.4 — Logging for post-market monitoring

AgentMint Feature Mapping
Evidence export (zip) Offline-verifiable package for regulatory submission.
VERIFY.sh + verify_sigs.py Independent verification without AgentMint installation.
SIEM-compatible fields Standard field names for integration with monitoring infrastructure.

Field Reference

Quick lookup: which receipt field maps to which framework requirement.

Receipt Field SOC 2 NIST AI RMF HIPAA EU AI Act Art. 12
id CC7.2 MEASURE 2.3 164.312(b) 12.1
plan_id CC7.2 MAP 1.1 164.312(b) 12.2
agent CC6.1 MAP 1.1 164.312(d) 12.1
action CC6.1 MAP 1.1 164.312(a) 12.1
in_policy CC7.2 MEASURE 2.3 164.312(b) 12.1
policy_reason CC7.2 MANAGE 3.1 164.312(b) 12.3
signature CC8.1 GOVERN 1.1 164.312(c) 12.2
previous_receipt_hash CC7.2 MEASURE 2.3 164.312(c) 12.2
evidence_hash CC7.2 MEASURE 2.3 164.312(c) 12.3
key_id CC6.1 GOVERN 1.1 164.312(d) 12.2
original_approver CC6.1 GOVERN 1.1 164.312(d) 12.2
policy_hash CC8.1 GOVERN 1.1 — 12.4
output_hash CC7.2 MEASURE 2.3 164.312(c) 12.3
reasoning_hash — MAP 1.1 — 12.3
session_id CC7.2 MEASURE 2.3 164.312(b) 12.1
session_trajectory CC7.2 MEASURE 2.3 — 12.4