Assume the Model is Compromised
Input defenses will never catch every hostile prompt, so this paper argues for the alternative: build systems where a compromised model can't do damage. A practical guide to sandboxing, isolation, and bounding the blast radius of agent runs, with evidence to prove it.
There Is No Firewall for a Sentence
You cannot secure a language model the way you secure software: it reads instructions and untrusted data as one stream, and nothing reliably tells them apart. Here is what building LuumenAI taught us about the one layer where you still have control.