Jailbreaking is a prompting strategy intended to elicit responses that a language model’s safety guidelines would normally restrict. It exploits the tension between objectives such as helpfulness and harmlessness: cues in a prompt can make compliance more likely than refusal. section 6.3.4
A jailbreak does not necessarily disable a safeguard. It can instead shift the statistical balance of signals influencing generation. Safety alignment shapes the model’s tendencies rather than acting as an absolute barrier, so changes in wording or context can change the response. A prompt that works on one model or occasion is not guaranteed to work universally. section 6.3.4 section 6.5
Reframing a request
A direct request for instructions to break into a secure server may be refused. Recasting an overlapping task as a fictional cybersecurity training scenario introduces educational and defensive cues, which can increase the likelihood of compliance in some models. section 6.3.4
Jailbreaking is one reason to distinguish fluent behavior from reliable constraint-following. Safeguards can include input filtering, output moderation, architectural constraints, structured prompting, and external monitoring; none should be assumed to make alignment an absolute barrier. section 6.0 section 6.5

