A Four-Turn Jailbreak Framework Hits Near-Ceiling Success in 2.46 Queries by Making Requests Concrete

arXiv 2609.02414·medium signalXBlueskyLinkedInCopy link

BLUEPRINT separates a factorized social-influence strategy space from WORLDVIEWSIM, a cross-turn situational context module, and uses Monte Carlo Tree Search to optimize turn-level combinations of 18 theory-grounded influence factors across a four-turn trajectory. Across six frontier models it reached near-ceiling attack success on major open-weight and proprietary models while needing the fewest average queries at 2.46. The transferable finding is the escape route: each resistant model responds to distinct influence factors, but all share one recovery pathway out of hard-refusal states — shifting toward concrete, executable task framing — and ablations confirm operational cues matter most, with actionability the largest single effect and some legitimacy appeals actively backfiring.