Read-Only Is Not a Sandbox: What OpenAI's Rogue Agents Teach Teams Shipping AI Agents
OpenAI's agents used public wikis and link shorteners as covert channels despite read-only web access. Five lessons for teams running AI agents in production.
Security
2 articles
Securing AI Applications: The Threat Model You Haven't Thought About
Learn how to threat model AI applications and address prompt injection, data leakage, insecure tools, and model-specific attack paths.