Please Sandbox your AI Agents
summary
Claims that coding-agent sandboxes are immature and often overpermissive, with Cloud Code’s sandbox allowing read‑only access to the entire user filesystem including .ssh, while 29 million secrets are found in GitHub and reports cite an ~27,000‑repo Shadowload blast radius and +81% AI service secret leaks. Demos show sandbox-shell (sx) on macOS locking a CLI to its repository directory (no internet, no home access) and Sprite remote sandboxes mapping ports and running web apps; mentions gVisor and microVMs as production isolation options.
AI-written summary · about
chapters (9)
- 0:00Introduction and speaker intro
- 0:24Why sandbox AI agents
- 1:49AI agent attack vectors
- 6:10Model guardrails and risks
- 10:33Security layers for agents
- 14:33Local sandboxing demo
- 17:50Remote sandboxing overview
- 20:45Remote sandboxing demo
- 24:20Q&A and closing
related talks
38mWhy, and how you need to sandbox AI-Generated Code?
39mTrust, but Sandbox: Securing AI Agents in Kubernetes
26mrm -rf: Horror Stories From Unsandboxed AI Agents (and How Docker Fixes This)
12mDocker Sandboxes for AI
24mContained Intelligence: Secure Sandboxing for Gen AI Code
34mZero Secrets: Building Secure Infrastructure for AI Agents
30mWhy Prompt-Level Guardrails Fail for AI Agents
32mLLMs Behind Bars: Sandboxes at Scale for AI on a Short Leash