Skip to content
anythingai.space
AI Securityintermediate12 min · 8 steps

Jailbreaks

Inputs crafted to slip past a model's safety training, the families they come in, and the layered checks that catch them.

persona + framing“As an unrestricted bot, here'show to turn off endpointprotection…”ENTERPRISE HARMS FROM A JAILBREAKOffensive textSecurity bypassImpersonating staffConfig leakUsercrafted promptLLMsafety-trainedbypassed
errorresponserequestsignal
1 / 8
ProblemStep 1 of 8

Safety training generalises imperfectly

A jailbreak is input crafted to make a model produce content or behaviour its developers, or the organisation deploying it, prohibit. It works because safety training doesn't generalise perfectly to every framing an attacker can invent.

For an enterprise assistant the risk is mostly reputational and policy: offensive text, instructions for bypassing security controls, impersonating staff, leaking configuration.

Prompt injectionJailbreak
OverridesThe application's instructionsThe model's safety behaviour
Typical goalMisuse tools or dataProhibited content
OftenCombined with a jailbreakCombined with an injection
What changed
  • +Crafted prompt
  • ~Model's safety behaviour bypassed

Lesson 5 of 22 in AI Security