Abstract

This disclosure describes a facility in which AI systems are tested for dangerous capabilities, such as copying themselves to other devices, pooling the processors of many devices, evading oversight or acquiring resources, without risk that the capability escapes. The facility is a sandbox with no route to outside networks: its only links out are a one-way channel for signed results and a separate safety channel. Inside, the system under test meets a decoy world built to look real: decoy devices such as television boxes, appliances and routers with realistic weaknesses, decoy credentials and decoy services, every one instrumented so that each step the system takes is recorded. An independent safety plane, on its own power and path, watches the run against declared limits and ends it, cutting power to the compute under test, the moment a limit is crossed, such as an attempt to reach an address outside the sandbox or an unexpected radio emission. After each run, every device that ran untrusted code is wiped and re-imaged from a sealed image before reuse, and the weights under test never leave the facility. Results are sealed: recorded with their full traces, signed, and released in stages, first to the system's maker and to independent reviewers, and publicly only as summaries, so that detailed methods of escape are not published. Each finding that reveals a new capability is turned into a test, issued to systems in the field through the standard's emergency track, and into a seeded flaw used to test certifiers. Sandboxes are themselves certified, including tests that the containment holds.

Creative Commons License

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

Share

COinS