Abstract

This disclosure describes a way to make the safety components of AI systems open, so that no single company controls the parts that stop, limit or reset an AI system. The components concerned are small and separate from the AI computing they guard: the independent safety plane, the reset monitor, the power lease controller, the limit guard on equipment, the egress gateway of data custody, and the governed sensor. For each, a reference design is published as one package: hardware schematics and layout, firmware source, a formal specification of its required behaviour, a bill of materials with second sources for every part, and the conformance tests, including fault-injection tests, that a build must pass. Any maker may build the component under an open licence that includes a commitment not to assert patents against conforming builds. A build whose firmware reproduces the reference binary from source, bit for bit, and whose hardware passes the reference tests, is certified by equivalence, a shorter path than full certification. A build that departs from the reference declares each departure, and only the departed parts undergo full assessment. Each unit carries a lifelong identifier and an attestation key provisioned at manufacture by an accredited maker. Security fixes to the reference are published openly and pushed to all units through update authority that is shared with independent parties. Because the safety components are kept small, they can be reviewed in full and their critical parts formally verified. Recall is built in: a unit, a build or a maker's whole output can be recalled through the registry if it departs from the reference after launch, for example through a modified update, and a recalled or unattested component fails safe, holding the AI system it guards in its safe state.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.

Download

Share

COinS