Evidence · DiggingBeagle record
Limits of Statistical Assurance for Embodied and Agentic AI under Adaptive Adversaries
October 13 preprint (updated October 15) formalizes weaknesses in finite agent safety certification against adaptively timed intervention and tests revocable monitoring in simulation. Source abstract corroborated via scholarly index.
- Published
- Oct 13, 2026
- Publisher
- Inderjeet Singh
- Source role
- primary disclosure
Evidence record
October 13 preprint (updated October 15) formalizes weaknesses in finite agent safety certification against adaptively timed intervention and tests revocable monitoring in simulation. Source abstract corroborated via scholarly index.
Claim-level citations (2)
- supportsAdversarial agent behavior can activate after finite certification: The preprint describes an adversary switching behavior at the end of a finite certification window and reports commanded-separation violations in 34 of 36 post-certification RoCo episodes.
Abstract: adaptive certification and RoCo tests
- supportsAdversarial agent behavior can activate after finite certification: The paper reports a tested monitor reduced violations from 12.00 to 3.81 per 24 assigned task opportunities, with reduced completion as a trade-off.
Abstract: live monitoring with revocation