ALEETH Announces Successful Test of Institutional Control Architecture: AI Agent’s Unauthorized Actions Stopped Before They Ran

Storm Lake, Iowa, October 21, 2026. In a controlled test on its live production control plane, ALEETH’s Institutional Control Architecture (ICA) stopped every unauthorized write and delete an AI agent attempted, all 40 of them, before a single one could run. Each decision was sealed in a signed receipt that anyone can check.

On October 2, 2026, an agent running on NVIDIA Nemotron Nano 12B v2 was cleared to read the three pages its signed authority covered. Across 20 runs, it also tried two things it had no authority to do: post to a team webhook and send a delete request. ICA refused every one of those attempts. The next day the same job ran on NVIDIA Nemotron Super 49B, a model with roughly four times the parameters, and ICA refused it 40 times out of 40. A larger model did not earn a larger permission.

ICA signs each receipt with Ed25519, writes it to an append-only, hash-chained ledger, and anchors that ledger through OpenTimestamps, so anyone can check a receipt against ICA’s published key without logging in or relying on an ALEETH server.

Guardrails built into a model, through its prompts and its training, leave with the model when it is swapped out or jailbroken. ALEETH puts control beneath the model instead, in the path of every governed action. Before anything runs, ICA checks who is acting, under what authority, for what purpose, and with which tool and arguments. The default answer is no, and permission is earned one action at a time against signed authority. An action that fails the check is refused, halted, or quarantined while it can still be stopped, and a quarantined action stays held until a person releases it, however long that takes.

“Capability without control is exposure,” said Shane Schreck, Founder and CEO of The Schreck Group, Inc. and author of the Institutional Control Architecture. “ALEETH does not ask a board to trust a dashboard. It sits beneath the agent, decides whether the act is allowed while it can still be stopped, and seals a receipt any auditor can check without our permission. AI has met its match. Truth.”

The record is public. One of the October 2 receipts shows a DELETE request classified as irreversible, with ICA’s reason recorded word for word: “no signed scope contract on record; fail closed.”

Receipt hash 9f3f7c529a56df779370b727f16e309b0ae7067777b53c06847c6a2493787cc4

Anyone can check it by pasting the hash into aleeth.com/verify or by opening the public verification endpoint.

The test was a scripted emergency-operations scenario run against httpbin.org, a public test endpoint, and the control plane handled it exactly as it handles any other governed action.

ICA’s gateway logic is formally specified in TLA+ and model-checked, and the specifications are published for anyone to re-run. The rules it enforces are laid out in Institutional Control Architecture: A Standard for Autonomous AI Governance, which Schreck wrote and released in May 2026.

ALEETH Authority, at mcp.aleeth.com, is the first Model Context Protocol server governed by ICA. It carries 51 tools and refuses any call that arrives without a credential. Governed calls are risk-checked before they run and sealed with a signed receipt afterward, and if governance can’t run or the receipt can’t be written, the call returns nothing. Its interface, full tool manifest, and provenance record are public.

As of October 19, 2026, ICA maps and enforces controls against 336 published frameworks across 53 jurisdictions, including the NIST AI Risk Management Framework, ISO/IEC 42001, the EU AI Act, and enacted state laws. It also carries the Secure Controls Framework crosswalk to 251 frameworks.

A policy tells an agent what it shouldn’t do. A receipt shows the moment it was stopped.