Agent Zero Trust

Open-source repository intake for AI coding agents, with experimental access-check and repair tools.

A repo is no longer just code. It is an instruction environment. Before you open an unfamiliar repository in a coding agent, inspect the instructions, hooks, configuration and setup commands that could influence it.

AZT is a deterministic, offline scanner. It flags known suspicious patterns, reports inspection gaps, and keeps the target from silently choosing its own exceptions. No model, account, telemetry or runtime dependency is required.

Install and scan

Python 3.9+ on Linux or macOS. Installation downloads the package; scanning runs offline and does not execute the target's code. Native Windows is unsupported.

python3 -m venv .venv-azt
.venv-azt/bin/python -m pip --isolated install --index-url https://pypi.org/simple --no-deps agent-zero-trust
.venv-azt/bin/azt --version
.venv-azt/bin/azt scan /path/to/authorized-repository

Use a repository you are authorized to inspect. Add --json for scope, findings, manifest and policy provenance. Exit codes: 0 passes the selected threshold; 1 has findings that meet it; 2 means incomplete inspection or an error. A clean scan is not proof of safety. You do not need to install a hook.

See a real scan

This is captured output from the published 0.1.9 wheel, scanning inert synthetic text. It never fetched or ran the command. Selected lines:

FINDINGS: 1 HIGH, 1 MEDIUM
  [HIGH  ] net.pipe_shell  README.md:1
EXCEPTIONS: 1 target requests (not applied); 0 trusted suppressed findings
DECISION: deny

Run the tiny demo, read the transcript, or view the static frame.

What changed in 0.1.9

The scanner's own trust boundary needed scrutiny, too. Target .azt-ignore content no longer suppresses findings. Operator exceptions must come from an explicit external policy and name an exact rule, path, content hash and reason. Suppressed findings stay visible. Incomplete inspection cannot quietly pass.

An optional workflow gate now authenticates a receipt for a specific reviewed snapshot. Edits can invalidate that receipt and require operator re-admission. The hook and a key accessible to the same user do not contain hostile code.

Upgrade and migration · Release notes · Why I’m strengthening Agent Zero Trust

Optional: test one access change and its repair

The experimental AZT-FS-001 pack compares an explicit, supported subset of Compose JSON bind mounts. It identifies selected added read access and proposes a minimal, digest-bound configuration diff for review. Static comparison needs no Docker and does not execute the supplied configuration.

On supported native Linux Docker hosts, a bundled trusted probe tests the change using fresh synthetic resources, applies the proposed repair to its disposable test plan, and retests while a small coding task runs. It does not test your real credentials or launch a live coding agent. Docker supplies the isolation; AZT adds configuration interpretation, repair, orchestration and reviewable evidence.

The accepted record covers two configuration inputs for one case, three phases each. Both retained original fixtures and completed cleanup. A passing misconfigured phase means the test demonstrated intentional exposure, not that the unsafe configuration is approved for deployment. Denial is credited only with the positive control and legitimate-task checks.

Exact evidence and limitations · Offline comparison and optional Linux reproduction

Know the boundary

The scanner detects known shapes, not every prompt injection or cross-file intention. Inventoried formats are not necessarily fully parsed. The public known-miss ledger predates this update and remains part of the project. FS-001 is a selected trusted-probe check, not universal containment, a model evaluation or an independent security audit. Other execution-boundary scenarios remain untested. The legacy doctor command concerns a deferred general runtime proposal, not FS-001's Docker prerequisite check.

Coverage and known misses · Supported files · Threat model · Security reporting

Help make it useful

Try an authorized scan. Report a missed detection, false positive or confusing result with a minimal synthetic fixture and the command you ran. Reproduce the documented case before generalizing its result.

Contribute · Open an issue · Use the Action · Public principles

Created by Rafael (Ralph) Peña, with credit to contributors and upstream work. The original engine came from rulebench vet. MIT. Citation metadata.

Delegate work. Retain control.