2026-01-21 · Authensor

AI Agent Compliance Reporting: What Auditors Need

Compliance reporting for AI agents requires demonstrable evidence that every agent action was evaluated against a defined policy, that decisions were logged immutably, and that denied actions were actually blocked — not just flagged. SafeClaw by Authensor produces hash-chained audit logs that satisfy these requirements out of the box, giving auditors a verifiable record of every allow and deny decision with timestamps, action details, matched policy rules, and cryptographic integrity proof.

Quick Start

npx @authensor/safeclaw

What Auditors Look For

External auditors and internal compliance teams need answers to five questions:

What can the agent do? — The active policy file documents allowed actions
What did the agent actually do? — Audit logs record every action request and outcome
Were unauthorized actions blocked? — Deny entries prove enforcement, not just monitoring
Can the logs be tampered with? — Hash-chained entries make tampering detectable
How long are logs retained? — Configurable retention satisfies regulatory timelines

Generating Compliance Reports

Export audit logs in structured formats:

# JSON export for automated processing
npx @authensor/safeclaw audit export --format json --since "30 days" > report.json

CSV export for spreadsheet review
npx @authensor/safeclaw audit export --format csv --since "30 days" > report.csv

Filtered export for specific action types
npx @authensor/safeclaw audit export --filter action=shell.execute --format json

Each log entry contains:

{
  "timestamp": "2026-02-13T14:32:01.447Z",
  "action": "file.write",
  "target": "/app/src/config.ts",
  "effect": "allow",
  "matchedRule": "rule-003",
  "policy": "team-backend.yaml",
  "provider": "claude",
  "hash": "a1b2c3d4...",
  "previousHash": "e5f6g7h8...",
  "sessionId": "sess_abc123"
}

Mapping to Compliance Frameworks

SOC 2

SafeClaw audit logs map directly to SOC 2 Trust Service Criteria:

| SOC 2 Criterion | SafeClaw Evidence |
|---|---|
| CC6.1 — Logical access controls | Policy files defining permitted actions |
| CC6.3 — Role-based access | Per-role policy configurations |
| CC7.2 — System monitoring | Real-time audit log of all decisions |
| CC8.1 — Change management | Version-controlled policy files in Git |

GDPR

For GDPR Article 30 compliance, filter logs to data-processing actions:

npx @authensor/safeclaw audit export \
  --filter action=file.read,file.write \
  --filter path="/user-data/" \
  --format json

Internal Audit

Generate a summary report showing policy effectiveness:

npx @authensor/safeclaw audit summary --since "quarter"

Output includes: total actions evaluated, allow/deny ratio, most-triggered rules, and policy coverage metrics.

Policy Documentation as Evidence

Auditors need the policy itself as an artifact. SafeClaw policies are plain YAML committed to version control, which means:

Version history shows when rules were added, changed, or removed
Code review on policy changes demonstrates approval workflows
Branch protection prevents unauthorized policy modifications

version: "1.0" description: "Production policy — auditor-reviewed Q1 2026" lastReviewed: "2026-01-15" reviewedBy: "security-team" rules: - action: file.read path: "src/**" effect: allow reason: "Read source for code generation"

- action: "*" effect: deny reason: "Deny by default — baseline enforcement"

Hash Chain Verification

Auditors can independently verify log integrity:

npx @authensor/safeclaw audit verify --since "90 days"

This recomputes the SHA-256 hash chain and reports any breaks, confirming no entries were inserted, deleted, or modified after the fact.

Why SafeClaw

446 tests ensure audit log correctness and hash chain integrity
Deny-by-default gives auditors confidence that unapproved actions are blocked
Sub-millisecond evaluation means compliance enforcement does not degrade performance
Hash-chained audit trail satisfies tamper-evidence requirements across frameworks
Works with Claude AND OpenAI — unified reporting across LLM providers
MIT licensed — no vendor-specific compliance entanglements