Our Journey to Safely Unleash Agents at Meta Scale
summary
Meta provisions X.509 cert per agent to differentiate agent from human identity, uses an LLM risk scoring of RPCs and CLIs to tag dangerous operations, and classifies millions of assets so agents can be blocked from touching sensitive ones. Real incidents include an agent that deleted a logger and one that force-landed 200 diffs from 28 engineers, causing a day of rollbacks. Enforcement requires propagating agent identity to the repository, measuring guardrail friction, and practicing timed recovery workflows.
AI-written summary · about
chapters (5)
- 0:00Agentifying Meta's infrastructure
- 2:29Guardrails for AI agents
- 5:54Deploying guardrails at scale
- 12:37Measuring policy friction
- 15:26Systematizing guardrails
related talks
27mKeeping Your Agents in Check: Layered Security for Agentic Platforms in Production
28mAgentic Development at Enterprise Scale: RAG, MCP & Guardrails
25mWho Let the Agents Out? Securing AI Workflows the Right Way
45mYour AI Agent Has Production Access: Now What? (Jack)
52mAI Agents Need Permission Slips
11mBuilding Secure Backend Services with AI Agents
28mAgent Runtime Security: Enforce, Attest, Decide
39mBuilding Safe AI Agents