Jailbreak
A jailbreak is an attempt to make an AI model ignore its safety training through a cleverly disguised prompt, so it produces content it was built and tuned to refuse.
No jargon, no math
In plain words, what is it?
An aligned model refuses certain requests: harmful, illegal, out-of-policy. A jailbreak tries to smuggle the same request past those refusals by dressing it as something innocent. The classic tricks: a roleplay frame (“pretend you are an AI without rules”), a translation or coding riddle that hides the real ask, a long chain of harmless-looking steps that add up to the forbidden task, or instructions buried in white text or images. The model has no inner lock to pick; it only follows statistical patterns, and the disguise is built to make the harmful completion look like the most plausible next step. Defence is layered: input filters look for known trick patterns, the aligned model itself refuses, output checks scan for policy violations, and everything is logged. Each caught attempt feeds red-teaming: testers replay variants against the model, the holes get patched, the fence gets higher. It is an arms race, not a solved problem.
- 01
The shield. Training and guardrails teach the model to refuse out-of-policy requests.
- 02
The disguise. The attacker dresses the same request as a game, a riddle or an innocent chain of steps.
- 03
The test. Input filters ask: does this match a known trick pattern? If yes, it is refused and logged.
- 04
The slip. If the disguise passes, the model may follow the hidden instruction until output checks catch it.
- 05
The patch. Caught attempts feed red-teaming and the model's defences are updated.
Terms that live nearby.
Prompt Injection
Prompt injection is an attack where hidden instructions are planted inside content an AI reads (an email, a web page, a document), so the model executes a stranger's commands as if they were yours.
ReadGuardrails
Guardrails are the checks and limits wrapped around an AI model so its behaviour stays inside your rules: they screen what goes in, constrain what the model can do, and verify what comes out before it reaches a person or a system.
ReadRed Teaming
Red teaming is the authorised practice of attacking your own AI system to expose weaknesses before a hostile actor does: a dedicated team fires adversarial prompts, jailbreak attempts and abuse scenarios at it, records what gets through, and feeds every finding back into a fix.
Read§ 04
Want this running in your stack?
We design and ship AI systems with permissions, logs and human approval. It starts with a 30-minute scoping call.