Secure Agent Design for PMs: Attacks, Architecture and a Checklist

Your agent has your keys, your files and the internet. What happens when someone attacks it? In this session, Mahesh Yadav breaks down how AI agents get attacked, why common designs fail, and the seven principles for building secure agents. He then walks through the architecture of a real production agent that applies all seven, so you can see them working together. What we cover: → What makes something an agent: input, model, policy, reasoning and tools → Four attack surfaces: stolen keys and non-human identity, prompt injection (including through files and web pages), policy and memory poisoning, and malicious MCP tools → Why designs fail: hardening only the edges, static credentials, no intent checks on tool calls, and unmonitored sub-agents → The 7 principles: a sandboxed key vault, vetted MCP servers and skills, an AI gateway for ingress and egress, immutable logs, continuous scanning of the harness, red teaming, and human in the loop → Case study: a production agent architecture with separate UIs for chat, approvals and credentials, a VM per user, a container that is destroyed after each request, and one-time payment cards for purchases → A security checklist you can hand to your team → Three answers to give when someone asks how you secure your agents → Who owns agent security: PM, engineering or security → How to trade off security, user experience, latency and cost Key takeaways for AI PMs: → Never expose credentials to the agent runtime → Sandbox every customer request → Check what goes into and out of your agent with a gateway → Log everything where the agent cannot edit it → Write security requirements in your PRD before engineering starts CHAPTERS 0:00 Intro and ground rules 2:10 What makes something an agent 7:04 How to attack an agent 9:06 Prompt injection, policy poisoning and MCP attacks 16:10 Why current designs fail 24:32 The 7 principles of secure agent design 30:07 Case study: a production agent architecture 39:06 Tainted requests and safe purchases 45:51 The security checklist 49:04 Three answers for "how do you secure your agent?" 52:47 The rule of 3: file access, internet, code 56:36 Who owns security: PM, engineering or security? 1:00:32 Security vs UX, latency and cost [ADD LINK: security checklist doc or session page] Next cohort adds a security-focused week: build the agent, secure it, then attack it. #AIAgents #AgentSecurity #AIPM #AgenticAI #PromptInjection