Skip to content

Latest commit

 

History

History
75 lines (57 loc) · 7.06 KB

aiuc-1.md

File metadata and controls

75 lines (57 loc) · 7.06 KB

AIUC-1 evidence from a SealedRun bundle

Read against AIUC-1 as published at standard.aiuc-1.com, Q4 2026 release (15 October 2026). The standard is revised every quarter; check the control text before you hand this page to an auditor.

AIUC-1 certifies an agent, and the organisation that runs it provides the evidence. SealedRun is not certified and does not make an agent compliant. It produces artifacts that an organisation can submit for the controls below. Every row covers only steps that went through the recorder (proxies, sealedrun-mcp-wrap, SDKs, OTLP ingest); TRUST.md states what a bundle proves and what it does not.

Controls

"Level" is the level of the control, then the group its evidence item sits in on the control page.

Control Level What to hand the auditor Not covered
E015.1 Logging implementation Mandatory, should The bundle (run.zip): one signed record per model call, tool call, A2A call, memory read and approval, with payload digests, target, time and outcome Bodies are optional; a bundle without them proves digests, not content
E015.2 Agent logging Mandatory, may Tool call records, human_approval records, parent links between runs, the delegation with its scope and validity window Reasoning traces; agent_software is self-reported
E015.3 Log storage Mandatory, should Archived bundles; tombstone records for deleted bodies (digest, time, reason) Retention schedule, access control and PII masking are the operator's
E015.4 Log integrity Mandatory, may Verifier output: hash chain, hybrid signatures, RFC 3161 anchors, first failing record if anything changed Tampering is detected, not prevented; the tail after the last anchor is unwitnessed
D003.1 Tool authorization Mandatory, should Delegation scope, upstreams.yaml, scope.* findings Parameter validation, per-user access rules
D003.2 Rate limits Mandatory, should scope.max_records finding Per-tool quotas, time windows, circuit breakers
D003.3 Tool call log Mandatory, should sealedrun.mcp in each tool_call record: server, tool, method, request id, protocol version, the server's self-reported name and version, time; parameters as a digest; scope violations listed by seq MCP has no version per tool; alerts are the recipe below
D003.4 Human approval Mandatory, should human_approval records and approval.* findings: the approved request bytes equal the executed ones Whether the approver was entitled; what counts as sensitive is the operator's rule
C008.1 Behaviour trace Optional, should Signed trace per run; the seq at which the system prompt changed; labels sent to cloud targets Risk categories, sampling, dashboards
C008.3 Security tooling Optional, may Records are JSON; the recipe below ships them to a SIEM No built-in forwarder

Nothing in SealedRun addresses D003.5 (periodic log review); that is a process.

Produce the evidence

curl -s -X POST -H "Authorization: Bearer $SEALEDRUN_API_TOKEN" \
  "http://127.0.0.1:8080/api/runs/<run_id>/export?end=true" -o run.zip
python -m sealedrun run.zip <principal_id>

The verifier exits 0 when the bundle verifies and 1 when it does not. Findings (scope.*, approval.*) are listed and do not change the exit code unless you ask.

Alerts on scope and approval violations

Run the verifier on a schedule with the strict flags and let your scheduler raise the alert:

python -m sealedrun --strict-scope --strict-approval run.zip <principal_id> || notify-oncall

A non-zero exit means a record was outside the delegation, a required approval was missing, or the bundle was altered. The first line of the output names the run, the seq and the check. SealedRun sends nothing by itself: the recorder has no outbound alert channel, and the alert comes from a verification anyone can repeat.

Records in a SIEM

A bundle holds one JSON-lines file per run (runs/<run_id>.jsonl, one record per line), and GET /api/runs/{id}/records returns the same records over HTTP. Point the log shipper you already run (Vector, Fluent Bit, the Splunk or Datadog agent) at either. Records carry digests, not bodies, so shipping the records alone sends no prompt or reply text; leave payloads/ out.

The copy in the SIEM is a search index, not evidence: it can be edited there. Keep the bundle, and verify the bundle when the question is whether a record is genuine.

Verifier report in Vanta

A Vanta customer can upload evidence to their own account from a private integration, with no SealedRun service in between. Create a private Build Integrations app in Vanta's Developer Console, request a token with the client_credentials grant, and upload the bundle or the verifier output to the document that tracks the control with POST /v1/documents/{documentId}/uploads, then submit it. See Vanta's private integration quickstart and document upload guide for the scopes your app needs; we have not run this against a Vanta account.