Skip to main content
Discipline Overview

What is AI Forensics?

AI forensics is the discipline of reconstructing what an autonomous AI agent did, why, and on whose authority — with tamper-evident, cryptographically-signed evidence that an auditor can verify offline, with no dependency on the vendor that ran the agent.

The Category Distinction

Forensics is not runtime security

Runtime security tries to prevent bad agent behavior in the moment. Forensics assumes something will eventually go wrong and prepares the evidence to reconstruct it afterward. Both matter. They are different jobs.

AI Runtime SecurityAI Forensics
Prevent prompt injection in the momentReconstruct which inputs the agent saw and how policy evaluated them
Block a jailbreak attemptProduce evidence the block happened and the chain wasn't tampered with
Rate-limit a misbehaving agentAttribute the misbehavior to a specific agent identity and policy state
Detect anomalous tool callsReplay the tool-call sequence with cryptographic proof of order and content

The Discipline

Four questions every audit eventually asks

AI forensics organizes evidence around four questions. The first three are governance primitives; the fourth is what makes the discipline forensic rather than operational.

PillarThe auditor's questionEvidence required
ForensicsWhat happened, provably?Signed Case Files, DSSE session attestations and public Evidence Anchor checkpoints, checked offline with the open-source verifier.
PolicyWhat was it allowed to do?Signed mandates with spend ceilings and validity windows, a fail-closed gateway, and a signed receipt for each draw, allowed or denied.
ComplianceWas the record kept properly?Retention policies, legal holds, tombstone receipts for pruned records, and redaction receipts for two-person erasure. Export profiles for SOC 2, EU AI Act and NIST AI RMF.
IdentityWhich agent did it?A key per agent, rotation without downtime, revocation and shadow-agent detection, so every record names the agent that acted.

Trigger Events

When AI forensics is the right tool

You don't need forensics to deploy an agent. You need it the first time someone asks for an account of what an agent did and a screenshot of the dashboard isn't enough.

Regulatory review

A regulator asks how an autonomous agent reached a decision that affected a customer, a payment, or a clinical outcome. The answer has to be reconstructable, not asserted.

Incident response

An agent took an action with material impact — a refund, a trade, a deletion, a contract acceptance. Root-cause analysis needs every step from authentication to action, with timing intact.

Customer dispute

A customer contests an action an agent took on their behalf. Evidence has to attribute the decision to a specific agent identity, policy state, and input context.

Compliance demonstration

EU AI Act, NIST AI RMF and SOC 2 reviews increasingly ask for demonstrable decision traceability for AI systems. Periodic attestation is not enough: auditors want per-action records.

The Open Reference

There's a specification for this

AI forensics now rests on open standards. OCSF 1.9.0 carries the attestation object and record_integrity profile, so a security event can carry its own integrity evidence. AID-EMIT-1 specifies how an enforcement point emits those records: what they contain, which bytes are signed, and how they chain, so anyone with the public key can verify them offline. The original v1.0 reference draft is archived as the historical record.