Abstract
This disclosure describes a defense for data custody systems, and for other services, against attacks carried out by autonomous AI agents. Four measures work together. First, certified agent admission: the service may require every automated client to present a certified identity for the AI system and a portable credential showing its conformance, and to act under a runtime governor whose current level and policy are attested; anonymous automation can then be refused outright, and every admitted agent's requests are recorded against its identity. Second, machine-paced attack detection: an independent safety plane, on its own power, processor and channel, watches for the signature of automated intrusion, such as rapid, systematic probing, enumeration of interfaces and credentials, and request timing and breadth that no human operator produces, and on detection isolates the affected node by itself, without depending on the systems under attack. Third, stolen data is made worth less: per-deposit encryption, keys held apart from the data, and output contracts that stop raw data from leaving mean that even a successful intrusion yields little that can be used or sold. Fourth, canary tracing: depositors may place unique synthetic records, known only to themselves, among their deposits; these are watched for in models outside custody, in data markets and on leak sites, and a match identifies the deposit while the ledger identifies every access to it, narrowing the source of a leak. The measures protect services that adopt them; they cannot govern attack tools that do not.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Wang, Edward D. H., "SI Data Bank™: Defending Data Custody against AI-Driven Attacks with Certified Agent Admission, Machine-Paced Attack Detection and Canary Tracing of Stolen Data", Technical Disclosure Commons, (October 13, 2026)
https://www.tdcommons.org/dpubs_series/12170