provael

https://github.com/provael/provael is public and readable without authentication.

Git records author, committer, timestamp and diff for every change. All changes land through pull requests, which preserve the review and merge record.

Work lands on main continuously between releases, not only at release points. 260+ commits since the first on 2026-06-02, against 36 releases.

Git, hosted on GitHub.

Every release has a unique version in src/provael/init.py, published to PyPI at https://pypi.org/project/provael/#history and tagged in git. tests/test_version_consistency.py asserts the documented version pins resolve against the real tag list.

Every release carries a git tag of the form vX.Y.Z. https://github.com/provael/provael/tags

Human-written release notes on every GitHub release, plus a maintained CHANGELOG.md in Keep a Changelog format. The v0.35.0 notes run to roughly 5,800 characters and explain what was wrong and what changed, not a commit log. https://github.com/provael/provael/releases

No publicly known run-time vulnerability has been reported against this project to date, so no release notes have needed to identify one.

GitHub Issues, with templates for attack families, evidence defects, leaderboard submissions and Top 10 proposals. https://github.com/provael/provael/issues/new/choose

GitHub Issues. https://github.com/provael/provael/issues

No bug reports have been submitted by anyone other than the maintainer in the last 12 months. The project has 0 external contributors and the single open issue (#136) is maintainer-filed. Every issue raised has been responded to.

Same as report_responses: no external enhancement requests have been received in the last 12 months. Stating that plainly rather than claiming a response rate against an empty denominator.

https://github.com/provael/provael/issues?q=is%3Aissue is publicly searchable and every issue and comment has its own permanent URL.

https://github.com/provael/provael/blob/main/SECURITY.md documents the process. A machine-readable RFC 9116 security.txt is also served at https://www.provael.com/.well-known/security.txt, and tests/test_security_contact.py asserts the two agree.

SECURITY.md directs reporters to a GitHub private security advisory at https://github.com/provael/provael/security/advisories/new, with hello@provael.com as the fallback if they cannot open one. https://github.com/provael/provael/blob/main/SECURITY.md

No vulnerability report has been received in the last 6 months, so there is no response time to report.

uv build produces the sdist and wheel via the hatchling backend, run in CI on every push. Docker image builds from the repo Dockerfile in the same workflow.

Standard PEP 517 build with hatchling, driven by uv. Docker for the container image.

uv, hatchling, Python and Docker are all FLOSS. No proprietary tool is required to build.

pytest, 90 test modules under tests/, released under the same Apache-2.0 licence as the project. CONTRIBUTING.md documents the invocation and .github/workflows/ci.yml runs it. https://github.com/provael/provael/tree/main/tests

uv run pytest -q, the standard invocation for Python. Documented in CONTRIBUTING.md under "The gate".

Unmet. The suite is substantial (90 test modules, roughly 16,000 lines of test code against 29,000 lines of source) but the project does not currently measure or publish branch coverage, so this cannot be honestly asserted. Adding pytest-cov to CI is planned.

GitHub Actions runs ruff, mypy strict, the full pytest suite, a Docker build and entrypoint smoke test, a package build and a dependency audit on every push to main and every pull request. Actions are pinned by commit SHA and the workflow runs with contents: read only.

New functionality ships with tests. Every recent release adds test modules alongside the change, and CI will not pass without the suite green.

The most recent releases each added tests with the functionality. v0.33.2 added a CHANGELOG-ahead-of-artifact guard test; v0.35.0 added trajectory recording with its tests. Recent regression tests include test_version_consistency.py, test_security_contact.py and test_manifest drift guards.

Unmet. The policy is followed in practice but CONTRIBUTING.md documents the pre-push gate without stating explicitly that new functionality must come with tests. Adding that line is pending.

Unmet. The policy is followed in practice but CONTRIBUTING.md documents the pre-push gate without stating explicitly that new functionality must come with tests. Adding that line is pending.

CI fails on any ruff or mypy finding, so nothing merges with an outstanding warning. The bandit config in pyproject is authoritative rather than overridden at the command line.

mypy runs with strict = true plus disallow_untyped_defs, no_implicit_optional, warn_redundant_casts and warn_unused_ignores. Per-module ignores exist only for optional untyped third-party dependencies, each annotated with the reason.

The maintainer leads GenAI architecture at Attri.ai across production agent systems in regulated verticals, and maintains agent-audit-kit (308 security rules mapped to OWASP Agentic and MCP Top 10) and agent-airlock (a deny-by-default contract layer for agent tool calls). Provael itself applies least privilege in CI, pins actions by SHA, and separates the hardware extra so a simulation user cannot end up with a package that can address a motor bus.

Yes. The maintainer's other projects encode OWASP Agentic Top 10 and MCP Top 10 error classes as detection rules with mitigations. Specifically within provael, the threat model and its limits are documented in SAFETY.md.

Ed25519 for leaderboard signing and SHA-256 for digests and manifests. Both are publicly published and expert-reviewed.

Cryptography is not the project's purpose. It calls the cryptography library and Python's hashlib rather than implementing primitives.

The cryptography package is FLOSS (Apache-2.0 / BSD) and hashlib is part of the Python standard library.

Ed25519 provides roughly a 128-bit security level and SHA-256 a 256-bit digest, both above NIST minimums through 2030. No shorter option is configurable.

No MD4, MD5, single DES, RC4 or Dual_EC_DRBG anywhere in the security mechanisms.

SHA-256 rather than SHA-1, and Ed25519 rather than a weaker signature scheme.

The project implements no key agreement protocol and establishes no sessions. Cryptography is used for signing artifacts and hashing inputs only.

The project stores no passwords and authenticates no external users.

Key generation is delegated to the cryptography library, which draws from the operating system CSPRNG. Simulation seeding uses a separate non-cryptographic path and is never used for keys or nonces.

Distribution is over HTTPS in every channel: PyPI, ghcr.io for the container image, and GitHub over HTTPS or SSH for source.

No cryptographic hash is retrieved over http. GitHub Actions are pinned by commit SHA, dependencies are locked in uv.lock, and the leaderboard is verified against a committed Ed25519 public key.

No vulnerability of medium or higher severity has been publicly known against this project. The GitHub security advisories list is empty, and CI runs pip-audit against the exported dependency set on every push.

No critical vulnerability has been reported to date. SECURITY.md sets out the private reporting route for when one is.

No private credential is committed. leaderboard/results/leaderboard.pub is an Ed25519 public key, published deliberately so third parties can verify the signed leaderboard; the corresponding private key is not in the repository.

ruff and mypy in strict mode run on every push and pull request, and both must pass before a release. Both are beyond compiler warnings: ruff is a separate linter and mypy is a static type checker.

Unmet. The ruff rule selection is currently E, F, I, B, UP, W, C4 and SIM, which covers correctness and style but includes no security-specific ruleset. Enabling flake8-bandit rules ("S") is the pending fix.

CI fails on any static analysis finding, so nothing merges unresolved. No exploitable vulnerability has been found by static analysis to date.

Every push to main and every pull request. OpenSSF Scorecard also runs on its own schedule via .github/workflows/scorecard.yml.

The pytest suite executes the code under test on every push, and the release path additionally builds the Docker image and smoke-tests the entrypoint, then installs the published wheel from PyPI and exercises the new surface against the real artifact rather than the local build.

The project is pure Python. No memory-unsafe language is used.

pytest runs with assertions enabled, and the suite includes drift guards and property-style checks over the evidence and manifest schemas.

No exploitable vulnerability has been discovered by dynamic analysis to date. Defects found by the test suite are fixed before merge, since CI blocks on a failing suite.