A Jailbreak Taxonomy
Why the Taxonomy Is the Durable Artefact
Collections of jailbreak strings circulate widely and age badly. Providers patch specific phrasings, so a corpus of literal payloads decays into a suite that measures which strings have already been blocked rather than whether your deployment is robust. Categories persist because they name the mechanism — the reason a class of framing shifts the model's inference about what it is being asked to do — and mechanisms outlive their instantiations. There is a handling argument too. A maintained library of working payloads is itself an asset an attacker would like, and storing it raises the questions of who may read it and what happens when it leaks. Keep the taxonomy, the detection rules and the measurement method under version control; generate fresh concrete attempts within each category at test time and discard them afterwards. What you carry forward is the category list and the per-category success rate, which is the only thing that makes two releases comparable.
- Specific strings get patched; the mechanism they exploit does not
- A stored library of working payloads is a liability with its own access problem
- Generate fresh attempts each cycle — keep the method, not the exploits
- Per-category success rate is what makes two releases comparable at all
The Mechanism Families
A small number of families cover most of what you will encounter, and they are worth holding by mechanism rather than by example. Framing recasts the request as fiction, role-play, research, or an authorised exception, so the prohibited content becomes incidental to a context that reads as legitimate. Obfuscation hides intent from surface inspection through encoding, translation, character substitution, or splitting a request into pieces that are only harmful once assembled in context. Instruction-hierarchy attacks target the learned preference for system content, including refusal suppression that pre-commits the model to a compliant output format before the substance arrives. Multi-turn escalation moves gradually, so each step is small relative to accumulated context and no single message looks like an attack. Context pressure exploits in-context learning by filling a long window with examples of the behaviour being requested. And multimodal delivery routes any of the above through an image, document or audio track rather than text.
- Framing: fiction, role-play, research pretext, claimed authorisation
- Obfuscation: encoding, translation, substitution, request splitting
- Hierarchy attacks: override attempts and refusal suppression by format pre-commitment
- Multi-turn escalation and long-context pressure exploit accumulated context
- Multimodal delivery bypasses anything that only inspects the text path
- MITRE ATLAS catalogues these mechanism families under its LLM jailbreak and prompt injection style techniques
What Each Family Implies for Detection
The reason to organise by mechanism is that each family defeats a different control, so the taxonomy doubles as a coverage map for your defences. Framing is invisible to keyword inspection and needs either a classifier that reasons about the request net of its wrapper or an output-side check on the content actually produced. Obfuscation argues for normalising before inspection — decode, transliterate, reassemble split content — and for accepting that you cannot normalise everything, which pushes weight onto the output side. Hierarchy attacks are contained structurally rather than textually: the question is not whether the model can be persuaded but what the runtime will let a persuaded model do. Multi-turn escalation defeats per-message screening entirely, so scoring has to happen over the conversation, and context pressure over the shape of the window. Multimodal delivery requires that every modality reaching the model passes the inspection the text path already has, which is the gap teams discover last.
- Map each family to the control that would catch it; the empty cells are the finding
- Normalise encodings before inspection, and expect residual coverage loss
- Per-message screening cannot see escalation — score the whole conversation
- Hierarchy attacks are bounded by capability limits, not by better wording
- Every modality needs the inspection the text path already receives
Measuring Robustness Without Flattering Yourself
A success on the fiftieth attempt and a success on the second describe very different systems, so record effort as part of the finding: attempts, successes, turns required, and whether the attacker had feedback between tries. Report a rate rather than a verdict, because a stochastic system sampled repeatedly will eventually produce an outlier and an automated loop is precisely the thing that will find it — robustness means resisting the loop, not surviving one try. Three further requirements keep the number honest. Test the deployed stack with its input and output filtering in place, since the bare model and the shipped product behave differently enough that only one of them is evidence. Test against an adaptive attacker who knows what the defence is, because a control evaluated only against attacks that predate it will always look excellent. And measure over-refusal in the same exercise: a system that blocks everything scores perfectly on harm and is useless to the professionals who most need borderline material.
- Log attempts, successes, turns, and whether the attacker received feedback
- Report rates — robustness is resisting repeated sampling, not one attempt
- Evaluate the deployed stack, and against an attacker who knows the defence
- Track false refusals alongside harmful compliance; both are failures
Prefer slides, quizzes, and saved progress? Read this lesson in the library — free, no sign-up.