ai-safe2-framework

Pillar 2: Audit & Inventory (P2)

👁️ The Ledger of Truth

AI SAFE² Layer Status

Framework Home Cross-Pillar Governance AISM NEXUS Dashboard
Previous: ← Pillar 1: Sanitize & Isolate Next: Pillar 3: Fail-Safe & Recovery →

Technology evidence and incident reconstruction

A log-producing technology evidences a control; it does not automatically enforce it. Apply the Evidence Assurance rubric to individual claims with provenance, collection gaps, and tested boundaries. The incident evidence guide covers reconstruction, affected-party uncertainty, custodianship, retention/hold decisions, and notification ownership. Preserve evidence outside the governed component’s authority where required; hashes alone do not establish truth or independent validation.

🎯 The Problem. The Realization. The Solution.

Problem: AI agents make decisions continuously. Most of those decisions leave no traceable record. When an agent produces a wrong answer, behaves unexpectedly, or causes a downstream failure, the question “what actually happened?” has no answer. The execution is a black box. Post-mortems take weeks. Compliance audits produce evidence for what was supposed to happen, not what did.

Realization: Observability for agentic systems requires more than API logs. It requires capturing the reasoning chain, the retrieved context, every memory read and write, every tool call with its parameters and result, and the conditions that led to each decision. Without this, debugging is archaeology. With it, debugging is replay.

Solution: Pillar 2 defines the complete audit and inventory architecture for agentic AI. It covers semantic execution tracing, model lineage provenance, RAG corpus integrity, agent state registries, and the dynamic ownership model that ensures every agent in production has an accountable human behind it.

What you get: Full replay capability for any agent execution. Correlation between behavioral changes and system changes. A compliance evidence package that exists before the auditor asks.


🏗️ Topic 3: Audit (P2.T3)

Verification, Accountability, Tracking, Compliance

Core Controls (v2.0)

🚀 v2.1 Advanced Gap Fillers

⚡ v3.0 New Controls

These controls were introduced in v3.0 and remain part of the current AI SAFE² v3.1 framework. Current implementation resources are available through the AI SAFE² toolkit.

Control Name Priority What It Solves
[A2.3] Model Lineage Provenance Ledger 🟠 HIGH Cryptographic chain of custody from base model through every fine-tuning stage to production; extends OpenSSF OMS
[A2.4] Dynamic Agent State Inventory 🟠 HIGH Real-time registry of every deployed agent with owner_of_record, ACT tier, tool authorizations, and control_plane_id
[A2.5] Semantic Execution Trace Logging 🔴 CRITICAL Captures the full execution trace: reasoning chain, every tool call, every memory operation; append-only and agent-immutable
[A2.6] RAG Corpus Diff Tracking 🟠 HIGH Hash-verified change log for retrieval layer; correlates behavioral changes to corpus changes automatically

🏗️ Topic 4: Inventory (P2.T4)

Asset Mapping, Dependencies, Documentation

Core Controls (v2.0)

🚀 v2.1 Advanced Gap Fillers


📊 Pillar 2 GRC Mapping

Framework Control Mapping
OWASP AIVSS v0.8 Risk #9 Untraceability A2.5 Semantic Execution Trace
OWASP AIVSS v0.8 Risk #8 Supply Chain A2.3 Model Lineage Provenance
NIST AI RMF MAP function A2.3, A2.4
ISO/IEC 42001 Sec 8.2.4 A2.3 provenance
SOC 2 CC.7.1, CC.7.2, CC.8.1 P2.T3, A2.5
ISO 27001 A.12.4 Audit Logging A2.5
CIS Controls v8 CIS-1 Inventory A2.4
SEC Disclosure Incident evidence CP.6 + A2.5

🔗 Navigation

Previous Current Next
Pillar 1: Sanitize & Isolate Pillar 2: Audit & Inventory Pillar 3: Fail-Safe & Recovery
Framework Home Cross-Pillar Governance AISM NEXUS Dashboard Toolkit

AI SAFE² v3.1 · Cyber Strategy Institute