From Alignment to Access Control: A Unified View of GenAI...
summary
Policy definitions vary across security engineering and AI research and that fragmentation breaks enforcement. The talk maps policies onto a concreteness vs enforceability quadrant, shows incidents (database deletion, sycophantic model drift) caused by relying on pretty please policies, and argues concrete rules like role or dataset access must use deterministic enforcement while fuzzy norms need model-steering. It proposes a donut model of residual stochastic holes and a runnable set of fixes: single control plane, standardized agent error codes, enforcement hooks, and conflict resolution for auditability and recovery.
AI-written summary · about
chapters (9)
- 0:00GenAI policy challenges
- 2:47Definitions of GenAI policy
- 6:18Policy enforceability model
- 8:28Pretty please enforcement issues
- 11:15Deterministic enforcement
- 12:41Stack enforcement model
- 14:48Model alignment for security
- 15:49Unified policy control plane
- 17:55Summary
related talks
30mRedefining Access Control: Scaling Policy as Code for Humans and AI Agents
33mGoverning the Ungovernable: Security and Compliance for AI Agents in Open Source Projec...
08mWhy AI Governance Matters More Than Ever
30mAgentic AI in Action: From Framework to Deployment
15mWhy Security Matters: The Risks of Agentic AI and How to Mitigate Them
01h 12mOut of Alignment Conflicting Trends in the Governance of Advanced Artificial ..
30mFine-Grained Authorization: The Missing Piece in Agentic AI Security
27mThinking Ahead About AI Security and Privacy