Your AI Has Too Much Access: Authorization for Coding Agents
summary
Claims that coding agents need deterministic, fine-grained authorization and shows an open source stack that implements it: SpiceDB (Zanzibar-style graph checks, 10M checks/sec scale) plus SpiceBox (local macOS sandbox, proxy egress, hook server) that translates English policies via Claude but enforces deterministic tuple/graph permissions. Provides concrete numbers — 17% false negative on auto-permission mode and 100% of apps tested with broken access control — three enforcement layers, five permission types, and temporary grants (about 15 minutes) with a human approval step.
AI-written summary · about
chapters (16)
- 0:00Intro and Stakes
- 0:49Coding Agents New Normal
- 1:26OWASP and Risk Stats
- 2:53Deterministic Enforcement
- 3:52Authorization Models Primer
- 6:35Zanzibar and ReBAC
- 8:07ReBAC Graph Explained
- 12:35SpiceDB Overview
- 14:07SpiceBox for Agents
- 15:59Policy Workflow Steps
- 17:48Three Enforcement Layers
- 21:06Permission Types and Approvals
- 23:45Limits and Caveats
- 25:21Live Demo Walkthrough
- 29:02Why Not settings json
- 29:27Final Takeaways and Wrap
related talks
55mYour AI Agents Are Overprivileged: The Case for Fine-Grained Authorization
33mHow to Prevent AI Agents from Accessing Unauthorized Data
30mYour AI Agents Are Overprivileged: The Case for Fine-Grained Authorization
16mAccess Granted with AuthZed
30mRedefining Access Control: Scaling Policy as Code for Humans and AI Agents
45mYour AI Agent Has Production Access: Now What? (Jack)
24mAI Governance for developers: A practical framework for deploying autonomous coding agents safely…
25mPlease Sandbox your AI Agents