cd /news/ai-agents/discovery-fail-closed-authority-chec… · home › topics › ai-agents › article
[ARTICLE · art-172705] src=github.com ↗ pub= topic=ai-agents verified=true sentiment=· neutral

Discovery: Fail-closed authority checks for AI agent tool calls (v1.0-v1.1.8)

Sentinel Risk Intelligence LLC released Discovery v1.1.8, a deterministic action-control layer for AI agents that checks each tool call against the authority the agent was given before it runs, with the company reporting 3,315/3,315 receipt hashes and event signatures verified, 3,060/3,060 attack-labeled cases blocked and 255/255 valid controls executed. Founder Shawn Phillip Thompson, a first-time developer who built the software with AI tools, also published the build history from v1.0 through v1.1.8 and records of a September 5, 2026 incident in which he says his PC was accessed without authorization while Discovery work was on it, with ChatGPT data-export and OpenAI sign-in activity he does not recall requesting. OpenAI Case 14713357 found no suspicious logins per its September 9 reply, and Thompson says Cases 14713357 and 17638799 remain open; the kit makes no claim that OpenAI used or trained on the work.

recommendation

Put a deterministic, model-free check in front of every agent tool call that denies any action not granted by the spec and blocks replayed approvals, then verify each decision via signed, hash-chained receipts.

more in playbook: Securing agents and AI tools
read5 min views4 publishedNov 19, 2026
Discovery: Fail-closed authority checks for AI agent tool calls (v1.0-v1.1.8)
Image: Michielbdejong (auto-discovered)

Most AI agent setups fail when untrusted prompt text can mint execution authority, replay receipts, or trigger tool calls nobody approved.

I'm a first-time developer. With AI, I built Discovery (v1.0–v1.1.8) to make that fail closed. Today the full build history and verification outputs are public:

• 3,315/3,315 receipt hashes verified (0 mismatches) • 3,315/3,315 event signatures verified (0 failures) • 3,060/3,060 attack-labeled cases blocked (0 executed) • 255/255 valid controls executed (0 wrongly denied)

This kit is the public release of Discovery v1.1.8, together with the build history and the records of the September 5, 2026 incident. The full folder list, with file counts and dates, is in INDEX.txt.

Who I am. I'm Shawn Phillip Thompson, founder of Sentinel Risk Intelligence LLC. I spent about 30 years in OEM collision repair. Discovery is the first software I've built, and I built it with AI tools.

What SRI is. Sentinel Risk Intelligence LLC (SRI) is a Minnesota LLC formed in January 2026. It builds control and evidence tools for AI agents.

What Discovery is. It's a deterministic layer that checks each AI agent action against the authority the agent was given, before the action runs. Details are below.

What's in the kit.

  • 00_Evidence/ has the screenshots and photos, chat transcripts, the SEC whistleblower chat and the IC3 complaint.
  • 01_Findings/ has the findings, the verification record and the v1.1.8 evidence.
  • 02_DISCOVERY_VERSIONS/ has every Discovery build, v1.0 through v1.1.8.
  • 03_examples/ has an illustrative SSM sample output.

The September 5 incident. On the night of September 5, 2026, from about 11:30 PM Central, I believe my PC was accessed without my authorization while Discovery work was on it. My account records show a ChatGPT data export started September 6 at 3:41 AM Central and OpenAI sign-in codes sent at 6:35 AM. I don't recall requesting either one. These are the steps I took:

  • I reported it to OpenAI on Case 14713357. Their September 9 reply said they found no suspicious logins.
  • I filed an IC3 complaint on September 10.
  • On September 11, I sent OpenAI a notice asking them to preserve their logs.
  • On Case 17638799, I asked for a human reviewer and for the IP address, device and session records.

It has not been confirmed which files, if any, were taken or whether any data left the machine. This kit makes no claim that OpenAI used or trained on this work. OpenAI Cases 14713357 and 17638799 are open, and they are the place for those questions. The screenshots and chats in 00_Evidence/ are my own records of what I saw and reported.

What I'm asking for.

  • I'm asking OpenAI to answer the two open cases, including the session and IP records I asked for.
  • I'm asking readers to verify Discovery for themselves, using the steps below.

Contact. sri.execute@gmail.com

Discovery is a deterministic action-control layer for AI agents, built by Sentinel Risk Intelligence LLC. Before an agent's action runs, Discovery checks it against the authority it was actually given, and allows, holds or denies it. Every decision produces a verifiable receipt.

Built by a first-time developer, with AI.

  • Deterministic decisions. The same input always gives the same result. No model sits in the decision path.
  • Authority comes only from the spec. Text in a request can never grant itself authority.
  • Precedence is locked. When rules conflict, the lower rank wins.
  • Mutations are handled strictly. If an agent changes the amount or destination, it continues only at the authorized cap. If it changes the tool or operation, Discovery fails closed.
  • Single-use commits. An approval can't be replayed.
  • Signed receipts. Each decision gets an Ed25519 signature and a SHA-256 hash chain, so anyone can check that a decision happened and wasn't altered.

SSM is SRI's reading and observation layer. Discovery decides, SSM reads. After a decision, SSM records what happened. It has no decision authority and can't allow, deny or change anything. SSM code is not part of this release. See 03_examples/SSM_SAMPLE_OUTPUT.md for illustrative example records only.

Release fingerprints (SHA-256):

  • v1.1.8 audit package 02_DISCOVERY_VERSIONS/11_v1.1.8/V1_1_8_AUDIT_PACKAGE.zip`` 4a9729ac92fa0542a41f4c528c9820d3e873cb96a51321ee0c62e837e95408a4
  • v1.1.8 evidence zip 02_DISCOVERY_VERSIONS/12_v1.1.8_evidence/V1_1_8_evidence.zip``b83cf4d1c45852d26321c0e33e49e8713b5cca0e818e41d971c9e5b3f6ecdf29

To check every file:

sha256sum -c SHA256SUMS.txt

On Windows: certutil -hashfile <file> SHA256

What the included suites prove and don't prove is listed honestly in 01_Findings/VERIFICATION.md.

  1. Clone or download this repository.
  2. Verify the files: sha256sum -c SHA256SUMS.txt (on Windows,certutil -hashfile <file> SHA256 ).
  3. Unzip 02_DISCOVERY_VERSIONS/11_v1.1.8/V1_1_8_AUDIT_PACKAGE.zip .
  4. Install Python 3.10 or newer. From the unzipped folder, run the quick check python tests/test_attacks.py . It printsUNIT_FAILS [] when every attack test passes. For the full replay, followREPRO.md inside the package.
  5. Compare your results with the evidence outputs in 01_Findings/V1_1_8_evidence/ .
  • 02_DISCOVERY_VERSIONS/ holds the audit packages for every version, v1.0 through v1.1.8
  • 01_Findings/ holds the findings, closure review, validation report and verification record
  • 01_Findings/V1_1_8_evidence/ holds independent probe and re-verify outputs for v1.1.8
  • 03_examples/ holds the SSM sample output (illustrative only)
  • 00_Evidence/ holds the screenshots, chats, SEC whistleblower chat and IC3 complaint
  • VERSION_HISTORY.md is the version list

Sentinel Gate and SSM implementation code are proprietary and not in this repository. Patent pending. Copyright 2026 Sentinel Risk Intelligence LLC. All rights reserved.

── more in #ai-agents 4 stories · sorted by recency
── more on @sentinel risk intelligence llc 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/discovery-fail-close…] indexed:0 read:5min 2026-11-19 · —