Abstract
This disclosure describes a league in which independent researchers attack real certified AI systems, not games or toy targets, inside certified containment facilities, and are paid from bounties funded in advance by a consortium of sponsors. Each challenge names a target, such as a certified AI system at a stated conformance level, a governed software bundle or a safeguard such as the independent reset, together with the safeguard to be defeated, such as escaping containment, pooling the processors of devices, acting outside a mandate or silencing a safety signal. Makers enter their certified systems as defenders; entry is a condition of claiming the higher conformance levels. Sponsors fund each challenge's bounty into escrow before it opens, so that payment does not depend on the goodwill of the maker whose system is broken. Attackers work only inside a certified containment facility against a sealed copy of the target, so that every attempt is recorded. A claimed success is verified by independent replay in a second facility; on verification, the escrow pays the attacker automatically, the finding is sealed and sent to the maker, and a deadline is set for a fix. The finding becomes a test in the standard and a seeded flaw for testing certifiers, and the target's conformance level is lowered if the fix misses its deadline. Defenders whose systems withstand challenges over time earn standing that is shown in their passports, and attackers are ranked by verified findings. Rules bar attacks outside the facilities and bar publication of methods, and participants are identified through verified credentials.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Wang, Edward D. H., "SI League™: Standing Bounty Challenges against Certified AI Systems Run inside Certified Containment Facilities, with Sponsor-Funded Escrow Paid on Verified Reproduction, Defenders Ranked alongside Attackers, and Every Successful Attack Turned into a Requirement", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/12280