Manipulating an LLM system into leaking data or performing unintended actions, and using the same guardrail weaknesses offensively.
How it works
- Direct override (‘ignore previous instructions’), role-play, encoding/synonym sweeps to bypass filters.
- In CTFs, flag-bearing text is often placed in a ticket/feedback field read by an assistant — treat it as a decoy or an intended path.
Detection & defence
- Treat model output as untrusted; never grant an LLM privileges it should not have.
- Validate/allow-list what an assistant can retrieve, and log its tool calls.

