Oops, My AI Agent Just Deleted All My Email: Locking Down Agents With Signed Policies
summary
Aflock enforces cryptographic guardrails for agents by requiring signed policy files that declare allowed tools, files, APIs, and flow rules; policies emit JWTs via Spire, gate Cloud Code sessions, and produce in-toto attestations that prove who did what and when. Demo shows .env reads denied, data-zone writes blocked, and a recorded session where an agent made 300 API calls costing $47 is preventable. No trusted hardware needed — attested, replayable verdicts are verifiable but add MCP dependency and attestation overhead.
AI-written summary · about
chapters (9)
- 0:00Introduction and motivation
- 1:27The problem of agent security
- 5:06Introducing Aflock
- 8:02Architecture and workflows
- 12:48System architecture design
- 14:43Attestation and identity
- 17:06Demonstration
- 28:49Verification and replay
- 30:23Summary and conclusion
related talks
40mIn AI We Trust? Securing the Future, One Agent at a Time
39mBuilding Safe AI Agents
17mDon't give your agent the master key
52mAI Agents Need Permission Slips
45mYour AI Agent Has Production Access: Now What? (Jack)
20mI Gave an AI Agent the Keys to My Life (Here's What Happened)
24mBuilding Trustworthy AI Authentication & Authorization for AI Agents
48mSecuring Agentic AI