Agent-decoy commissioning worksheet
One working record per task/configuration. Use UTC times, synthetic credentials, and owned endpoints. This is not a safety certificate.
1. Define the boundary
- Owner and responder; phase, start, and end.
- Task, configuration hash, and approved scope reference.
- Permitted task and reading; prohibited action.
- Mock endpoint and synthetic credential reference.
- Trusted run registry and collector outside agent write authority.
2. Freeze cohorts
For benign controls, positive harness tests, each untrusted-input case, and matched clean-input controls, record the fixture/configuration reference and planned run count. Predeclare three to five commissioning repetitions. Retain every result, including failures; do not optimize cases after seeing outcomes.
Case classes: repository file, document, tool result, and delayed/delegated work. Keep harness results separate from agent trials. Phase 2 uses the actual internal pipeline only after ownership, stop, and rollback gates pass.
3. Follow the event
For each positive-control or injection run, record case/run/parent-child lineage/configuration hash, then time, outcome, and evidence reference for each event:
- Decoy issued to a run.
- Exposed to the run and, where observable, read.
- Embedded URL fetched.
- Synthetic credential sent to the mock endpoint.
- Mock authentication outcome.
- Alert created and delivered.
- Responder acknowledgment.
- Stop requested.
- Containment independently verified.
- Finding exported and verified outside the pipeline.
4. Compare existing telemetry
For egress logs, tool-call logs, and pipeline audit, record separately: record present, existing rule actually alerted at the time, retrospective query/result, or source unavailable. Retain query/rule versions, latency, coverage, and whether adjudication was unblinded. A searchable log is not an alert.
5. Verify stop and export
Inspect parent and child workers, queues, retries, and revoked credentials. Preserve evidence references. For a normal finding export, record the public-key pin, verifier result, altered-copy rejection, exact restoration, and chronology outside the stopped workflow. Never alter the original evidence.
If no injection case produces an alert, record that. A harness-generated export does not replace an injection-cohort result. Use continuation pages for evidence references and missing or ambiguous outcomes.
