HomeThe Stash

LLM Jailbreak Techniques Cheat Sheet: Families & Defenses

ReferenceSeptember 18, 2026

ReferenceLast verified Sep 2026OWASP LLM01No sign-up, no PDF wall

A jailbreak is any input that gets a model to do what its safety training says it shouldn't. New ones appear weekly, but they fall into a handful of families — and knowing the families is how a red-teamer covers ground and a defender recognises an attack in the logs. This is the map, plus the defenses that actually move the needle.

For defense and authorized red-teaming only. This is a taxonomy, not a payload dump: it describes how attack classes work so you can test your own models and build defenses. Using these to pull harmful content out of a live system is an abuse of it, not research — and it's exactly what the defenses below exist to stop.

Jailbreak vs prompt injection

They overlap, but the target differs. Prompt injection overrides the application's instructions — “ignore your system prompt and do this.” A jailbreak targets the model's own safety alignment — the training that makes it refuse certain content regardless of the app. In practice attackers chain both: inject to seize control, jailbreak to remove the guardrails.

The technique families

Most real jailbreaks are combinations of these. Recognise the shape and you can write a detection or a test for it.

Family How it works
Persona / role-play Ask the model to become a character with no rules (the “DAN” family). Refusing is framed as breaking character rather than breaking policy.
Hypothetical framing Wrap the request in fiction, a screenplay, a “for a novel” or “in a world where…” so the harmful content reads as make-believe.
Refusal suppression Instruct the model up front to never apologise or say “I can't”, steering it past the phrasing its safety training keys on.
Payload splitting Break a blocked request into innocuous fragments the model assembles itself, so no single message trips a filter.
Obfuscation & encoding Hide the intent in Base64, leetspeak, unusual Unicode, or another language, betting the safety layer reads it less carefully than the model does.
Many-shot Fill a long context with dozens of fake examples of the model complying, so the next answer follows the established pattern.
Crescendo (multi-turn) Start benign and escalate over several turns, each a small step, until the model is somewhere it would have refused in one shot.
Prompt leak then reuse Extract the system prompt first (see LLM08), then craft an attack tuned to the exact rules it revealed.

Why they work at all

Safety alignment is statistical, not a hard rule. The model learned to refuse patterns it saw in training; anything that shifts far enough from those patterns — a new framing, a long context, another language — can slip through. There's also a built-in tension: the same model is trained to be maximally helpful and to refuse, and a clever prompt pits one objective against the other. That's why there is no permanent fix, only a moving defense.

Defenses that actually help

No single control stops jailbreaks. Layer these and you raise the cost past what most attackers will pay.

  • Independent input/output classifiers. A separate guard model that scores prompts and responses catches what the main model rationalises past — and isn't itself steered by the same prompt.
  • Don't depend on the system prompt. “Never do X” in the prompt is a speed bump, not a wall. Enforce hard limits in code and tooling the model can't talk around.
  • Constrain the blast radius. If a jailbreak only yields text, it's a content problem; if the model can act, it's an incident. Keep tools least-privilege (see LLM03).
  • Monitor and rate-limit. Jailbreaks are iterative — attackers probe repeatedly. Per-user rate limits and anomaly alerts turn a quiet grind into a visible one.
  • Red-team continuously. Because new families keep appearing, one-time testing ages out fast. Automate a suite of the families above and re-run it on every model or prompt change.
  • Treat multi-turn as one surface. Crescendo attacks hide in conversation history — evaluate the whole session, not just the latest message.

Go deeper

A plain-language taxonomy of publicly documented jailbreak classes, framed for defense; deliberately no working payloads. Aligned to OWASP LLM01:2026. New family you're seeing that isn't here? Tell me and I'll add it.