AI Agents Need Permission Slips
summary
A concrete blueprint for safe agents: an agent at Replet in July 2025 executes a drop database and wipes ~1,190 customer accounts, prompting mandates for cryptographically secure session IDs in the November 2025 MCP spec update. Use MCP server with OAuth 2.1 style short‑lived tokens (1–15 minute access, refresh ~24 hours), enforce attribute-based access control, score risk (auto approve <70, require 2FA >90), put humans in the loop for destructive ops, isolate dev/staging/prod credentials, and log every decision for fast audit.
AI-written summary · about
chapters (12)
- 0:00AI agent security risks
- 3:22Real-world AI agent disaster at Replet
- 10:28Deployment questions checklist
- 15:24Static authorization failures
- 18:36Dynamic authorization lifecycle
- 28:09Runtime defense in depth
- 31:57Token and session handling
- 33:42Model Context Protocol security
- 35:26Human approval gates necessity
- 43:11Five-step deployment framework
- 49:58Production deployment checklist
- 51:43Summary and call to action
related talks
23mIdentity Best Practices to Secure & Scale Agentic AI / MCP Deployments
26mWho Let the Agent In? Securing MCP Servers in Production
29mMCP Security: Keep Your AI Agents from Spilling the Tea
01h 22mIdentity for AI Agents
19mHow to Secure Agents using OAuth
15mWhy Security Matters: The Risks of Agentic AI and How to Mitigate Them
40mIn AI We Trust? Securing the Future, One Agent at a Time
45mAPIs are the best protection for your AI strategy -