Launching today

GROUND IMMUNE Guard
Separate what the AI decides from what it's allowed to do
2 followers
Separate what the AI decides from what it's allowed to do
2 followers
Giving an AI a job is not the same as giving it unlimited authority to act. GROUND IMMUNE Guard checks each operation an agent proposes against explicit permissions and evidence, and returns ADMITTED, DENIED, or UNRESOLVED (exit 0/3/4 in LIMITED_BLOCKING). Default SHADOW mode only observes. LIMITED_BLOCKING uses a mock Broker. Technical Preview, not production-proven. Apache-2.0, Node.js 22+. Try it from the GitHub Release tarball; it is not on the npm registry.








Free
Launch Team

AWS Startup AdvisorPersonalized AWS guidance that meets you where you’re already building
Promoted
Hi Product Hunt, I'm Shunsaku Yasutake. I built GROUND IMMUNE Guard at ORIMUSU Inc.
Giving an AI a job is not the same as giving it unlimited authority to act. A flawed plan, a prompt injection, or a plain bug can still trigger operations that run with whatever permissions the agent already has. I wanted a local way to separate what the AI decides from what it's allowed to do.
Guard checks each operation an agent proposes against explicit permissions and evidence, and returns one of three verdicts:
- ADMITTED (exit 0 in LIMITED_BLOCKING): in allowed scope, required evidence present.
- DENIED (exit 3): outside the allowed scope (OUT_OF_SCOPE).
- UNRESOLVED (exit 4): required evidence missing, a decision reused, or the target changed after admission. It is put on hold, not treated as an allow.
Who it's for: people who are already wiring agents to real operations and want an inspectable verdict before those operations run.
What you can try today (Technical Preview, v0.1.0-preview.4)
- No account or credentials. After download it runs offline. Node.js 22 or later.
- The default mode, SHADOW, only observes: it records the verdict and blocks nothing. In SHADOW, every verdict exits 0 unless you pass --fail-on-verdict.
- LIMITED_BLOCKING demonstrates blocking with a mock Broker. In the shipped example, an operation swapped after admission comes back UNRESOLVED / TARGET_CHANGED, and the Broker function runs 0 times.
- I ran the same release on Linux, Windows 11, and macOS (Apple Silicon); the documented outputs matched.
- Apache-2.0. The Release tarball includes the TypeScript source (SHA-256 checksums and an SBOM, not signed). The public GitHub repository holds docs, examples, and validation records; the TypeScript source is not in the repository tree.
What it is not: this is not a proof of production safety. It does not demonstrate credential isolation, there is no real Broker host or real write executor in the package, and it is not on the npm registry. The limits are in docs/LIMITATIONS.md.
I'd like to hear:
1. In your agent setup, where would a verdict like this sit, and what would the runtime need to honor it?
2. If you run the Quickstart, where does it lose you?
Site: https://orimusu.com/en/
Repo and Quickstart: https://github.com/ORIMUSU/ground-immune-guard
Release: https://github.com/ORIMUSU/ground-immune-guard/releases/tag/v0.1.0-preview.4
Security issues: security@orimusu.com, privately please, not in the comments. Pilots and product questions: immune@orimusu.com. Thanks for taking a look.