Coding agents run code by design, read untrusted content like issues and web pages, and often hold credentials. Prompt injection can turn that content into instructions, so you have to assume the model can be steered. The lethal trifecta of private data, untrusted content and an exfiltration path is a quick test for every setting you choose.
Read in this order
- ArticleThe lethal trifecta for AI agents (Simon Willison)Start here: the three-part test you will apply to every configuration choice.
- ArticlePractical security guidance for sandboxing agentic workflows (NVIDIA AI Red Team)Concrete controls from a red team for limiting what a steered agent can do.
- VideoVibe Check: Security Failures in AI-Assisted IDEs ([un]prompted 2026)Turns exploit patterns across coding tools into a hardening checklist. 25 min.