Abstract
This disclosure describes a way of testing and improving AI systems, and the AI agents that work on them, by running attack, defence and learning together and without pause, rather than as separate exercises months apart. A Red side attacks a target within a signed, expiring scope lease; a Blue side detects, defends against and recovers from those attacks; and a Purple side, led by a certified person, watches both, triages every finding, and turns each one into a verified fix, an executable test and, where it applies generally, a proposed update to a living standard. The three sides may include people and AI agents. To keep the exercise honest, the Red and Blue sides are kept apart: each has its own credentials, its own lease and its own runtime, with no shared memory, channel or training feed between them during an exercise, so that no participant, human or AI, can play both sides or quietly weaken its own attack or defence. The Purple side sees both sides but cannot act on targets. Every exercise action is labelled and signed, so that exercise traffic is never mistaken for a real attack and a real attack is never dismissed as an exercise. Exercises run first on targets the operator owns or controls, such as certified containment facilities, digital twins and pilot services, and only later on third-party systems under signed authorization. The tests produced by the Purple side join regression suites against which every new version of an AI system must pass before release, in staged waves; a system's own test results never certify it, and certification remains with an independent accredited body. Findings that apply widely are offered to the living standard and may become emergency clauses. The loop then runs again, so that each attack strengthens the defences of every system that adopts the resulting tests.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Wang, Edward D. H., "SI Purple Team™: Continuous, Simultaneous Red, Blue and Purple Teaming of AI Systems, with Red and Blue Sides Kept Apart by Separate Credentials, Leases and Runtimes, a Human-Led Purple Function that Turns Every Finding into a Verified Fix, an Executable Test and a Standard Update, and Releases of AI Systems Gated by the Resulting Test Suites", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/12644