liveBook · Manning

Jailbreaking

As explained in Machines that Think

A jailbreak is an input or interaction strategy intended to bypass a system's safety restrictions or behavioral constraints.

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

How Machines that Think covers it

How deep it goes
Explains how framing can shift statistical cues toward compliance rather than refusal, without detailing a safeguard implementation.
Examples
Contrasts a direct server-break-in request with a fictional cybersecurity training scenario framed around defense.
What you can do after
Evaluate how prompt framing can affect refusal and why safeguards are not absolute.
How it connects
Connects jailbreaking to safety alignment, probabilistic generation, input filtering, and output moderation.

Also explained in

Compare how 2 books explain jailbreaking →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Machines that Think. The section links point to the parts of the book it draws on.