railward

Version updated for https://github.com/Ourbando/railward to version v0.2.0.

• This action is used across all versions by ? repositories.

Action Type

This is a Composite action.

Go to the GitHub Marketplace to find the latest changes.

Action Summary

Railward is a deterministic guardrail that automatically tests and validates AI policies. It runs predefined attack scenarios against the policy to ensure it behaves as expected. The action provides signed, hash-chained logs of allowed, blocked, and leaked actions, allowing users to verify the policy’s compliance and detect potential vulnerabilities.

What’s Changed

railward v0.2.0: first public release.

A fail-closed, veto-only gate for untrusted AI coding agents that ships with its own adversary and a re-runnable signed proof. The agent proposes; a small pure function decides allow, deny, or ask; and every run signs a hash-chained proof of what was blocked. Flip one rule and the proof goes red.

• 41 attack classes, all refused by the example policy
• fail-open probe battery (a broken policy asks, never allows)
• Ed25519 signed, hash-chained proof; verify in the browser with nothing installed
• Claude Code PreToolUse hook, plus a CLI and a Python API
• offline, pure Python, MIT

pip install railward

• Ourbando
• GitHub Actions