Abstract
This disclosure describes two linked safeguards against the misuse of AI agents for attacks. First, AI agents built for legitimate security testing, such as penetration testing and red-team exercises, are governed so that they cannot be turned against systems whose owners have not agreed. Such a testing agent carries a portable credential, here called a passport, tying it to a certified operator; it runs only against targets named in an authorization signed by each target's owner and checked against a registry; it works under an expiring scope lease that names the targets, the permitted techniques, the dates and the limits, and refuses anything outside it; it labels and signs all of its traffic, so that defenders can tell an authorized test from an attack; it has a stop that does not depend on the agent's own software; and every action is recorded in a signed record shared with the target's owner. Copies of a testing agent that are stolen, modified or used outside their authorization are revoked through the registry, and the agent refuses to run without a valid passport and lease. Second, a real-time quarantine network lets certified services detect and share malicious agents as they act. When a certified service sees agent behaviour that matches declared attack patterns, such as many failed logins at machine speed, use of decoy credentials, unlabelled testing traffic or activity outside a lease, it publishes a signed report of indicators, without personal data, to the network. Reports from independent reporters are corroborated, and when a declared threshold is reached, every subscribing certified service blocks the indicated agent, infrastructure or credential within seconds. Blocks are graded by confidence, expire unless renewed by fresh evidence, and may be appealed by anyone wrongly blocked. Where the malicious agent presents a certified passport, that passport is suspended everywhere at once and the operator's certification is reviewed. When a service provider holding access to many customers is quarantined, every customer's grant to it is suspended at once, affected systems step down to their declared safe modes, and customers are told through a verified channel. Signed reports are kept as evidence and may be passed to the proper authorities by the operators concerned.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Wang, Edward D. H., "SI Quarantine™: A Real-Time Quarantine Network in which Certified Services Report Malicious AI Agents, Corroborate Reports and Isolate Them across the Network within Seconds, with Confidence Grades, Expiry and Appeal, together with Governed Security-Testing Agents that Run Only against Targets whose Owners Have Signed an Authorization, under an Expiring Scope Lease, with Labelled and Signed Test Traffic", Technical Disclosure Commons, (November 02, 2026)
https://www.tdcommons.org/dpubs_series/12618