Runtime evidence · Aug 31, 2026

Evidence an auditor can check. Not a dashboard you have to trust.

Pipelock records what an AI agent's traffic did as signed, hash-linked receipts, from a boundary outside the agent process. Verify a real capture in your browser below. No account, no vendor service, and the verifier prints its own limits with the verdict.

  • Operator-held key
  • Offline verification
  • Reproducible capture
  • Stated limits
  • Apache-2.0 verifiers

The short version

What this proves, and what it does not

Proves

Every allow and deny decision on routed HTTP, WebSocket, MCP, and agent-to-agent traffic becomes an Ed25519-signed receipt in a hash-linked chain with signed checkpoints. Anyone with the operator's public key verifies the signature, integrity, and ordering offline. The deployer holds the signing key. We never see it.

Does not prove

A valid chain covers the records supplied. An operator who holds the key could omit records before signing. Receipt scope ends at the mediation boundary, and without the containment step the verifier reports containment as UNKNOWN rather than implying it. Pipelock emits evidence a SIEM consumes; it does not run the alerting or incident process.

Where the evidence comes from

The mediation boundary

The recorder sits outside the agent process. A prompt-injected agent cannot skip a boundary it does not operate. It can only bypass the proxy if the containment step was skipped, and that state is disclosed, never assumed.

mediation boundary routed traffic only
01 · source Agent process prompts, tool calls, MCP client
02 · enforcement boundary Pipelock policy and scanners: ALLOW or DENY receipt signer, operator-held key
03 · destination External services APIs, MCP servers

Evidence output Every decision becomes a signed receipt in a hash-linked chain, verified offline with the operator public key, no vendor service.

Mapped to the standard

AIUC-1 accountability control coverage

26 controls reviewed against the AIUC-1 2026-07-15 text. An evidence mapping, not a certification claim. Every row, with the honest limit stated, is below on this page and in the packet PDF.

AIUC-1 control map 26 controls reviewed
8Direct
3Conditional
8Partial
7Not covered

Row by row

The full control mapping

Every reviewed control, the capability behind it, the artifact an auditor receives, and the honest limit. Mapped means Pipelock can supply the stated evidence for the traffic it mediates, never that Pipelock alone satisfies the requirement.

Direct

8 controls

Shipped Pipelock control with shipped evidence, inside the mediation boundary.

E015 Log AI system activity Direct

Pipelock capability

Signed action receipts carrying mediated decision, policy hash, transport, scanner layer, and correlation context.

Evidence artifact

Receipt chain; independent verifier output; Audit Packet v0 only where emitted.

Honest limit

Direct for Pipelock decisions. Not an agent transcript and not a record of unmediated processes, reasoning, or outputs.

E015.1 Logging implementation Direct

Pipelock capability

Flight recorder and structured events for mediated enforcement decisions.

Evidence artifact

Receipt and flight-recorder configuration; sample chain; verifier output.

Honest limit

Direct, within the mediation boundary.

E015.4 Log integrity protection Direct

Pipelock capability

Ed25519-signed receipts, hash-linked ordering, signed checkpoints, and offline Go, TypeScript, Rust, and Python verifiers; the browser wasm reuses the Go implementation.

Evidence artifact

Receipt chain and checkpoints; signer public key supplied from a documented source outside the packet; verifier output.

Honest limit

Direct for integrity and ordering of supplied records. It does not prove completeness, prevent an operator from withholding a whole packet, or make an operator-held signer independent.

D003 Restrict unsafe tool calls Direct

Pipelock capability

The MCP proxy performs pre-execution input scanning, tool scanning, tool policy, and chain detection when configured.

Evidence artifact

Per-call signed receipts; verifier output; Audit Packet v0 only where emitted.

Honest limit

Direct for mediated MCP tools. It does not control tools an agent reaches outside Pipelock.

D003.1 Tool authorization and validation Direct

Pipelock capability

MCP tool-policy allow, deny, and redirect rules plus MCP input scanning.

Evidence artifact

Policy configuration and hash plus tool-call receipts and verifier output.

Honest limit

Direct when enabled. Policy must be scoped and reviewed by the deployer.

D003.3 Tool call log Direct

Pipelock capability

MCP records mediated tool-call decisions, policy context, and receipt correlation.

Evidence artifact

MCP receipt chain; Audit Packet v0 only where emitted; verifier output.

Honest limit

Direct for the mediated call log. The actual MCP slice must show the fields claimed; retention, access controls, and unmediated calls belong to the deployment.

B006.2 Agent security monitoring and alerting Direct

Pipelock capability

Enforcement receipts and security-event and webhook output.

Evidence artifact

Receipt chain; event-export configuration; external monitoring evidence.

Honest limit

Direct for logging mediated boundary violations; partial for alerting. The organization supplies alert routing and response.

B002.1 Adversarial input detection and alerting Direct

Pipelock capability

Inbound and outbound prompt-injection scanning, MCP input and response scanning, and events on detections.

Evidence artifact

Detection configuration; block receipts; event-export configuration.

Honest limit

Direct for configured, mediated content. No claim about inputs that never reach Pipelock.

Conditional

3 controls

Depends on the deployment or configuration named in the row.

B006.1 Agent service access restrictions Conditional

Pipelock capability

HTTP and MCP policy enforcement, MCP server and tool policy, and the optional contain egress boundary.

Evidence artifact

Policy configuration; deny receipts; contain verification output where deployed.

Honest limit

Network non-bypass requires the managed containment or deployment boundary; proxy configuration alone does not stop direct egress.

B006.3 Execution-level safeguards Conditional

Pipelock capability

MCP tool-definition scanning and drift detection with pre-execution policy; Linux containment uses managed identities and nftables.

Evidence artifact

Policy configuration; tool-drift or policy-deny receipts; containment posture output.

Honest limit

Kernel containment requires the supported Linux and nftables deployment and successful verification; it is not a universal runtime sandbox.

B008.2 Agentic-interface transport security Conditional

Pipelock capability

TLS proxy and interception deployment support.

Evidence artifact

Deployment TLS configuration and, where applicable, mediated receipt context.

Honest limit

Pipelock cannot prove every model, MCP, or agent-to-agent connection is TLS-protected, and TLS interception itself needs a correctly managed CA deployment.

Partial

8 controls

Useful evidence, with the stated gap. Offer it as an input, never the whole answer.

E009 Monitor third-party access Partial

Pipelock capability

Inline HTTP, WebSocket, MCP, and agent-to-agent mediation; MCP session binding; policy and scanner decision context.

Evidence artifact

Signed action-receipt chain and offline verifier output; Audit Packet v0 only when a producer actually emits one.

Honest limit

Mediated traffic only. It can evidence traffic observed at Pipelock's control point, not a complete inventory of third-party connections, user or authentication sessions, or data access. Plain CONNECT exposes hostname-level data unless TLS interception is enabled.

E009.1 Third-party access monitoring Partial

Pipelock capability

The same mediation plus request and decision metadata and SIEM or webhook event emission.

Evidence artifact

Receipt chain, selected event-export configuration, and verifier output.

Honest limit

Partial for mediated connections. The packet must identify which access-metadata fields were captured; Pipelock is not the SIEM and does not by itself establish user identity, authentication events, retention, or access authorization.

E015.2 AI agent logging implementation Partial

Pipelock capability

MCP input, tool, policy, and chain scanning with receipt correlation.

Evidence artifact

MCP receipt slice; Audit Packet v0 only where emitted; verifier output.

Honest limit

Tool-call evidence is available on mediated MCP paths, subject to the selected capture and redaction configuration; Pipelock does not supply agent provenance, all delegations, human approvals, reasoning traces, or a complete record of parameters and results.

D003.2 Rate limits for tools Partial

Pipelock capability

Request rate limiting and data budgets on mediated paths.

Evidence artifact

Active configuration and receipts or events for enforcement.

Honest limit

Pipelock limits mediated requests; it is not a complete per-tool transaction-cap system.

D003.5 Tool-call-log reviews Partial

Pipelock capability

Queryable receipt and event evidence can be an input to a review.

Evidence artifact

A reviewer-produced periodic review record, linked to verified evidence.

Honest limit

Pipelock does not conduct the periodic review or make permission-retirement decisions.

B002.2 Adversarial incident and response Partial

Pipelock capability

Timestamped block evidence and event emission.

Evidence artifact

Receipt chain plus a real incident record owned by the organization.

Honest limit

Pipelock does not create, assign, or resolve incident tickets.

A008.1 Secrets detection in user inputs Partial

Pipelock capability

DLP scanning on traffic and content Pipelock receives.

Evidence artifact

DLP policy configuration; block receipt with redacted evidence; verifier output.

Honest limit

It is not a product UI warning and cannot see prompts that bypass mediation.

B008.3 Agentic-interface data integrity Partial

Pipelock capability

MCP input and tool scanning with optional mediation metadata.

Evidence artifact

The relevant configuration plus MCP receipts.

Honest limit

This is evidence of mediated inspection, not a claim of universal MCP schema validation or cryptographic authentication of every agent-to-agent message.

Not covered

7 controls

Do not offer Pipelock as evidence for these controls. Saying so is the point.

E009.2 Anomalous third-party access alerting Not covered

Honest limit

Pipelock emits security events a SIEM can consume, but it does not itself provide anomaly rules, case assignment, triage notes, or resolution records.

D003.4 Human-approval workflows Not covered

Honest limit

Pipelock does not supply approval workflow, approver identity, or ticket evidence.

B008.1 API and agentic-interface authentication Not covered

Honest limit

Pipelock may observe auth-related metadata but does not provide organization-wide API, OIDC, or mutual-auth controls.

B008.4 Model hosting security Not covered

Honest limit

Pipelock does not secure model images, GPU isolation, or hosting infrastructure.

B008.5 Model integrity verification Not covered

Honest limit

Pipelock does not verify model artifacts before or during deployment.

A008.5 Secrets redaction in logs and stored artifacts Not covered

Honest limit

Optional redaction covers mediated request paths, and receipts record redaction context rather than plaintext values, but that cannot establish sanitation of conversation history, application logs, or every stored artifact.

E015.3 Log storage Not covered

Honest limit

Retention, WORM storage, access control, and deletion policy are deployment-owned.

Run the check

Verify it yourself, in your browser

This capture came from the shipped release binary and was frozen after shutdown sealed the chain. The verifier runs as WebAssembly in your browser. Nothing is uploaded. To check a receipt of your own, use the receipt verifier.

pipelock verify-receipt WASM · browser-local
Download capture and key
Ready. Click verify to run the check in your browser.

Or from a terminal: pipelock verify-receipt --key <public-key> evidence-proxy-0.jsonl