When securing AI agents, have you asked whether a control makes the attack impossible or merely tedious? Title: Zero Trust for AI Agents URL: 🛡️ Claude's eBook lays out a Zero Trust framework for deploying autonomous AI agents in the enterprise. I could read only the first 8 pages. 🎯 Highlight 1: Impossible, not tedious Rate limits, non-standard ports, and SMS-based MFA degrade badly against agentic attackers with unlimited patience and near-zero cost per attempt. What survives: hardware-bound credentials, expiring tokens, and network paths that simply don't exist. When in doubt, remove a capability instead of throttling it. 💥 Highlight 2: Blast radius An agent with admin access to cloud infrastructure has an enormous blast radius, while read-only access to one database has a small one. "Design for breach" means assuming every agent's blast radius will eventually be tested. 🔑 Highlight 3: Least agency A term coined by OWASP that extends least privilege to each agent tool. A database tool gets read-only queries, an email summarizer gets no send or delete rights, an API gets minimal CRUD. Restrict what a tool can do, how often, and where. A ready-to-use yardstick for designing agent permissions. #AISecurity# #ZeroTrust#