Paper · arXiv
AI Safety: Not Optional, Not Later
Incidents show that AI safety failures often arise across multiple layers. We present a safety-by-design assurance architecture combining model-level supervision, such as Scientist AI, with system-level controls over scaffolds and harnesses, independent verification, monitoring, and evidence infrastructure, supported by governance for accountability and evidence interoperability.
| Authors | Qinghua Lu, Yoshua Bengio |
|---|---|
| Primary category | cs.SE |
| Published | 9 September 2026 |
| Last updated at source | 9 September 2026 |
Silytics shows public metadata from arXiv and links to the original. Information may be incomplete or out of date. Last refreshed 17 October 2026. Licenses and terms of the original apply to the work itself. Not financial, legal or professional advice.