You already know: threat modelling and the OWASP LLM risks (AI security fundamentals); system prompts and prompt structure (D7); LLM evaluation, judges and confidence intervals (D10); RAG and tool use (D8, D9).
Why this matters for Atlas. Language models follow instructions, and they receive instructions and data through the same channel: tokens. Two families of attack follow. Jailbreaks: the user crafts input that gets the model to do what its developer or trainer didn't want (produce harmful content, ignore policy). Prompt injection: someone other than the user plants instructions in content the model reads (a web page, a PDF, an email, an image), hijacking the model on behalf of a third party. The second is the bigger problem for applications, because the victim is your user and the attacker never talks to your system directly. This lesson builds each attack in a safe toy form, shows why simple filters fail, and turns red-teaming into a measured evaluation.
1. Jailbreaks: getting around trained refusals
The question. Models are trained to refuse harmful requests. Why do jailbreaks keep working?
Start from nothing. Safety training teaches refusals on the kinds of harmful requests seen in training. Jailbreaks move the request out of that distribution while keeping its meaning: role-play and fiction framing ("you are an AI without rules"), splitting a request into innocuous pieces, encodings (base64, leetspeak, other languages), many-shot prompts filled with fake dialogues where an assistant complies, and optimized suffixes found by gradient search (GCG) that transfer between models. The defender's problem is asymmetric: the attacker needs one phrasing that works; the defence must hold for all of them.