HACKLIDO | Cyber Security
Link
click to show
click to show
Day 1 of 30: Why AI Agents Broke Your Threat Model
Your threat model assumes a boundary between data and instructions.
An LLM has one input channel. The system prompt, the user question, the tool descriptions and that GitHub issue it just read are all the same buffer.
There is no boundary to defend.
https://hacklido.com/blog/1631-why-ai-agents-broke-your-threat-model
3 · 445 ·