{"paper":{"title":"Break the Brake, Not the Wheel: Untargeted Jailbreak via Entropy Maximization","license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","headline":"Maximizing entropy at high-entropy refusal tokens flips VLM outputs to harmful content without fixed targets and improves cross-model transfer.","cross_cats":["cs.AI"],"primary_cat":"cs.CV","authors_text":"Jing Zhang, Jinhong Ni, Mengqi He, Shu Zou, Weikang Li, Xin Shen, Xinyu Tian, Xuesong Li, Zhaoyuan Yang","submitted_at":"2026-05-11T15:59:02Z","abstract_excerpt":"Recent studies show that gradient-based universal image jailbreaks on vision-language models (VLMs) exhibit little or no cross-model transferability, casting doubt on the feasibility of transferable multimodal jailbreaks. We revisit this conclusion under a strictly untargeted threat model without enforcing a fixed prefix or response pattern. Our preliminary experiment reveals that refusal behavior concentrates at high-entropy tokens during autoregressive decoding, and non-refusal tokens already carry substantial probability mass among the top-ranked candidates before attack. Motivated by this "},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Our experimental results indicate that the limited transferability primarily stems from overly constrained optimization objectives.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"Refusal behavior concentrates at high-entropy tokens during autoregressive decoding, and non-refusal tokens already carry substantial probability mass among the top-ranked candidates before attack.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"UJEM-KL improves cross-model transferability of untargeted jailbreaks on vision-language models by maximizing entropy at decision tokens instead of forcing specific outputs.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Maximizing entropy at high-entropy refusal tokens flips VLM outputs to harmful content without fixed targets and improves cross-model transfer.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"fc5d00d99395f5bbb300cf1cfa64b688303d889a24af5dfbb3c1036bb80ba050"},"source":{"id":"2605.10764","kind":"arxiv","version":2},"verdict":{"id":"a3237278-d062-425b-986f-5d6978ec1dd0","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-12T03:32:55.001710Z","strongest_claim":"Our experimental results indicate that the limited transferability primarily stems from overly constrained optimization objectives.","one_line_summary":"UJEM-KL improves cross-model transferability of untargeted jailbreaks on vision-language models by maximizing entropy at decision tokens instead of forcing specific outputs.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"Refusal behavior concentrates at high-entropy tokens during autoregressive decoding, and non-refusal tokens already carry substantial probability mass among the top-ranked candidates before attack.","pith_extraction_headline":"Maximizing entropy at high-entropy refusal tokens flips VLM outputs to harmful content without fixed targets and improves cross-model transfer."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.10764/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"claim_evidence","ran_at":"2026-05-20T05:22:00.400417Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"ai_meta_artifact","ran_at":"2026-05-19T14:36:17.535678Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_title_agreement","ran_at":"2026-05-19T11:01:16.758894Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T09:00:03.992680Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"4eae519190e07d72913539afbad17157daf12591da49f26910a6c2bb73fd4daf"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}