In a landscape where AI agents often operate with loose permissions, Shawn Phillip Thompson, founder of Sentinel Risk Intelligence LLC, has released 'Discovery' (v1.0–v1.1.8), a deterministic action-control layer designed to fail closed when authority checks are ambiguous. Unlike typical agent frameworks that rely on model inference for decision-making, Discovery enforces strict precedence rules where text in a request can never grant itself execution authority. The release, hosted on GitHub, includes the full build history and verification outputs for a September 5, 2026, incident that prompted the tool’s development.
Zero-Mismatch Verification Data
The project’s credibility rests on extensive cryptographic verification. Thompson reports that 3,315 receipt hashes and 3,315 event signatures were verified with zero mismatches or failures. Furthermore, the system successfully blocked 3,060 attack-labeled cases with zero executions, while correctly allowing 255 valid controls to proceed without wrongful denial. The kit provides SHA-256 fingerprints for the v1.1.8 audit package and evidence zip, allowing independent developers to verify file integrity using standard tools like sha256sum or Windows certutil.
The September 5 Incident Context
Discovery’s development was catalyzed by a security scare on September 5, 2026. Thompson alleges his PC was accessed without authorization while work was in progress, citing a ChatGPT data export initiated at 3:41 AM Central and OpenAI sign-in codes sent at 6:35 AM—events he does not recall requesting. Although OpenAI’s Case 14713357 reply found no suspicious logins, Thompson filed an IC3 complaint and opened a second case (17638799) requesting session and IP records. The release explicitly makes no claim that OpenAI trained on this specific work, but the incident underscores the risks of unmonitored agent tool calls.
Deterministic Architecture and SSM Layer
The core philosophy of Discovery is determinism: the same input always yields the same result, with no model sitting in the decision path. If an agent attempts to mutate an action—such as changing a tool or operation—the system fails closed. Approvals are single-use and cannot be replayed, with each decision generating an Ed25519 signature and a SHA-256 hash chain. The release also references SSM (Sentinel State Monitor), a read-only observation layer that records events but holds no decision authority, though its implementation code remains proprietary.
Key Takeaways
- Fail-Closed Design: Untrusted prompt text cannot mint execution authority; strict precedence rules apply.
- Cryptographic Proof: All 3,315 receipts and signatures are verified with zero mismatches.
- Incident-Driven: Built after a suspected unauthorized access event on September 5, 2026.
- First-Time Dev: Thompson, with 30 years in OEM collision repair, built this tool using AI assistance.
- Open Verification: Public SHA-256 hashes allow anyone to audit the release integrity.
The Bottom Line
Thompson’s release is a crucial reminder that AI agent security shouldn't be an afterthought. While the 'first-time developer' angle is compelling, the rigorous zero-failure verification data proves that deterministic guardrails are not only possible but essential for production-grade agent autonomy.