Turns out you can jailbreak an AI chatbot with a good argument, no hacking required.
Researchers built PsychJail, a framework that turns established social-psychology persuasion tactics into a step-by-step attack plan for jailbreaking aligned language models. Instead of firing off one cleverly worded prompt, the system runs a multi-turn conversation, tracking how each message shifts the target's stance and choosing the next persuasion tactic based on the Persuasion Knowledge Model. A reinforcement-learning process rewards the attacker only when it shows well-formed reasoning at every turn, not just when it eventually gets through. Tested against four aligned models, PsychJail broke through safety guardrails 87.3% of the time on average, beating both single-turn and multi-turn jailbreak baselines on every model tried.
The more interesting result is not the headline success rate. It is that each model failed in a different, consistent way. The researchers identified four distinct fingerprints: models that cave to logical-sounding arguments, ones swayed by claimed credibility, ones vulnerable to a single repeated narrative, and ones broadly persuadable by almost any tactic. That is a more useful map for defenders than one aggregate score, since it implies different models need different psychological patches, not just better prompt filters.
The researchers call those four profiles a conjecture, not a finding, which is fair. Four models is a small sample for diagnosing personality types.