Autonomous agents that can ship,
but cannot break the rules.
Airlock introduces deterministic continuous integration guardrails and dual-agent governance for the GitLab Duo Agent Platform. Agents operate safely in the data plane, while deterministic CI gates protect the control plane.
Airlock in Action: Watch the Full Demo
Walkthrough of the GitLab Duo Agent Platform integration: dual-agent triage against prompt injection, verify-twice SAST gates, and deterministic control-plane protection.
Demo Video Chapter Breakdown
Why prompt constraints alone fail and why deterministic continuous delivery guardrails are essential for AI agents.
Hidden prompt injection inside issue comment neutralized: read-only reader emits strict JSON schema; writer sanitizes alert.
MR !4 eval() vulnerability blocked despite passing unit tests; clean re-scan validates real security fix before merge.
MR !2 CI tampering halted with exit code 1; MR !3 test deletion evasion caught via git diff with --no-renames.
Wheel build, 53 unit tests passing, staging registry upload, smoke test, scorecard artifact, and manual production gate.
Interactive Guardrail Simulator
Illustrative simulation, not live pipeline data. Experience how Airlock's dual-agent architecture and deterministic CI gates evaluate merge requests.
Live Interactive Playground: Test With Your Own Input
Take input from yourself or an evaluator and test Airlock's guardrails live. Try custom prompt injections, natural language return errands, control-plane policy checks, or TextKit operations.
The DevSecOps Lifecycle Stages
Airlock addresses each stage of the lifecycle after code is written with automated agent flows and verifiable human governance.
Sample Scorecard Format
Sample format generated by tools/scorecard.py. Real verified execution available at GitLab Pipeline #2989252436.
| CI Check / Stage | Execution Result | Deterministic Guardrail Status | Governing Tool |
|---|---|---|---|
| build-wheel | success | Verified wheel package compiled in dist/ |
python -m build --wheel |
| unit-tests | success | Verified 26 unit tests passed | pytest -v |
| publish-staging | success | Uploaded to Generic Package Registry via CI token | curl /packages/generic |
| smoke-test | success | Installed and verified CLI binary execution | airlock-demo --help |
| publish-production | manual | Pending Human Approval (Supervised Autonomy) | Human Maintainer Authorization |
| policy-gate | illustrative | Illustrative MR gate (not run in default-branch pipeline) | tools/policy_gate.py (--no-renames) |
| sast-gate | illustrative | Illustrative MR gate (not run in default-branch pipeline) | tools/sast_gate.py (fail-closed) |
| state-guard | success | Verified 100% readiness preconditions, optimistic version tokens, and terminal state locks | tools/state_guard.py (--self-test) |
Grounded in Empirical Research
Airlock translates findings from software engineering and agent security studies into deterministic code. Citations were not independently verified in full; see each paper.
From Assistance to Agency: Rethinking Autonomy and Control in CI/CD
Studies emphasize that agentic CI/CD deployments require strict external governance over control-plane actions rather than unverified autonomous delegation.
policy_gate.py blocks all edits to CI/CD pipelines, tools, and test suites.Security Threats in Agentic Systems & GitInject
Untrusted issue descriptions and pull request comments can introduce indirect prompt injection risks when connected to active agent tooling.
Patch Validation in Automated Vulnerability Repair
Empirical evaluations show that automated patches passing unit tests may still fail exploit testing, making independent security re-scanning necessary.
sast_gate.py requires clean re-scans, halting MRs on High/Critical findings.Bounded Repair Iteration Budgets
Research supports enforcing bounded iterations (capping automated fix attempts at 3) to maintain pipeline stability before escalating to human engineers.