{"paper":{"title":"Jailbreak Scaling Laws for Large Language Models: Polynomial-Exponential Crossover","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Strong adversarial prompt injections turn slow polynomial growth of jailbreak success into exponential growth with more inference samples.","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Annesya Banerjee, Cengiz Pehlevan, Indranil Halder","submitted_at":"2026-03-11T21:48:03Z","abstract_excerpt":"Adversarial attacks can reliably steer safety-aligned large language models toward unsafe behavior. Empirically, we find that adversarial prompt-injection attacks can amplify attack success rate from the slow polynomial growth observed without injection to exponential growth with the number of inference-time samples. We first identify a minimal statistical mechanism for these two regimes by giving a small set of assumptions on the distribution of safe generation across contexts under which both scaling laws follow. To explain this phenomenon further, we propose a theoretical generative model o"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"strong adversarial prompt-injection attacks can amplify attack success rate from the slow polynomial growth observed without injection to exponential growth with the number of inference-time samples","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"a small set of assumptions on the distribution of safe generation across contexts under which both scaling laws follow","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Strong adversarial prompt injections shift jailbreak success scaling from polynomial to exponential with more inference samples, derived from a spin-glass generative model of LLM outputs.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Strong adversarial prompt injections turn slow polynomial growth of jailbreak success into exponential growth with more inference samples.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"62c11c0d1a70bcc0f3c9f75dcc2ee35722dead7734d613a721ab458319785aca"},"source":{"id":"2603.11331","kind":"arxiv","version":3},"verdict":{"id":"46c5e45c-d2dd-4f7c-baf8-c678a02e2eb5","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T12:37:31.020882Z","strongest_claim":"strong adversarial prompt-injection attacks can amplify attack success rate from the slow polynomial growth observed without injection to exponential growth with the number of inference-time samples","one_line_summary":"Strong adversarial prompt injections shift jailbreak success scaling from polynomial to exponential with more inference samples, derived from a spin-glass generative model of LLM outputs.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"a small set of assumptions on the distribution of safe generation across contexts under which both scaling laws follow","pith_extraction_headline":"Strong adversarial prompt injections turn slow polynomial growth of jailbreak success into exponential growth with more inference samples."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2603.11331/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":2,"snapshot_sha256":"1fe285041d1e5885d3400c8cde56fc8851d3c4936f4553eea7c537da583f13f4"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}