What changes in your IR playbook
For any AI-involved incident: (1) identify the time window. (2) query receipts /v1/receipts?from=&to=. (3) filter by suspect agents / users / projects. (4) replay each receipt to see the actual tool calls + arguments + responses. (5) build the timeline + the harm assessment from the replay. (6) verify signatures on every receipt in the timeline to confirm chain of custody hasn't been broken. The receipt IS the evidence.
Why signature verification matters
Receipts are signed at issuance. Any post-hoc tampering invalidates the signature. So your IR chain is cryptographic: if the receipt verifies, you can present it as legally admissible digital evidence. If it doesn't, you have an attack signal (someone tried to tamper). Standard Ed25519 verification; takes a few lines of code; runs offline.
Anomaly detection categories
Cost spikes (runaway loops / stolen credential). Off-hours activity (compromised laptop). Atypical model use (someone's using a more capable model for unauthorised tasks). Receipt-count surges (data exfiltration via AI prompting). Signature failures (config drift OR active tampering). Each category maps to a known attack pattern.
Integration with SIEM + SOAR
Anomaly alerts ship as JSON to your webhook. Splunk / Datadog / Chronicle ingest. SOAR playbooks: cost spike → freeze the agent's API key. Off-hours → page the on-call. Signature failure → quarantine the laptop, force re-install of the MCP server. Standard SOC patterns; the receipt feed is the data source.
Forensic acquisition
/v1/receipts/export?from=&to=&format=evidence produces a signed forensic bundle: all matching receipts + the signature verification proof + the agent registry snapshot + a chain-of-custody manifest. Hand this to forensics or external counsel; standard evidence-acquisition workflow.
What we don't do
We don't prevent attacks (we're audit + alerting, not policy enforcement). We don't catch DLP violations directly (we catch AI-mediated ones via anomaly). We don't replace your EDR. Position us as the AI-specific layer in your defence-in-depth; not as a replacement for anything you already have.