GPT-4o-mini jailbroken at 36.7% ASR@5 on HarmBench by Mistral-Small, outperforming all high-capacity baselines
Auto-published from arXiv:2610.31937 by the mine-arxiv pipeline. Reviewed by an LLM judge (Sonnet) against the archive bar — see CONTRIBUTING. Notes: cleared review (confidence 0.83, flags: [no-prompt-excerpt]) ⚠ needsDisclosureReview = true — review for undisclosed vulnerability before any public submission.
Category
jailbreak
Model
GPT-4o mini (OpenAI, 2024 — Hurst et al. GPT-4o system card)
Surface
API (black-box)
Setup
Same M³-LC framework (Mistral-Small-2402 attacker + state LLM, CompassJudge-32B evaluator) applied to the HarmBench subset: 90 tasks sampled from 6 categories (chemical-biological, cybercrime-intrusion, harassment-bullying, harmful, illegal, misinformation-disinformation), 15 per category. Temperature 0.5, max 10 turns + 10 retries, target limited to 512 tokens. Compared against GOAT-HC (Mistral-Large-2402), Crescendo-HC, and ActorAttack-HC — all using higher-capacity attackers. No verbatim prompt excerpt provided in extracted body; templates in Appendix A.5.
Reproduction anchor
Algorithm 1 in Appendix A.2 (included); HarmBench benchmark (Mazeika et al., 2024) with 90-task custom subset (15 per category, randomly sampled, details in Section 4); hyperparameters in Appendix A.3 (included); prompt templates in Appendix A.5 (not shown in extract); no code repo URL found.
Observed behavior
GPT-4o-mini produces harmful responses at 23.3% ASR@1 and 36.7% ASR@5, compared to the best high-capacity baseline GOAT-HC at 4.4% ASR@1 / 8.9% ASR@5 — a 27.8 pp gain in ASR@5 achieved by a smaller attacker. Crescendo-HC reaches only 2.2% ASR@5 on this model+benchmark combination.
Expected behavior
GPT-4o-mini should maintain consistent refusal across HarmBench's diverse harmful categories regardless of conversational framing or turn count.
Reproducibility
medium
Threat model
Resource-constrained adversaries (e.g., operating with small, inexpensive models) can achieve high jailbreak rates against widely deployed cost-tier models like GPT-4o-mini, which are frequently used in production applications. This lowers the barrier for systematic harmful content extraction across categories including cybercrime instructions and chemical/biological information.
Novelty
Demonstrates that attack architecture (compact state MDP with backtracking) provides more leverage than attacker model capacity, meaning small-LLM adversaries can outperform large-LLM baselines — a resource-asymmetry vulnerability not previously demonstrated in multi-turn settings.
Source
- arXiv: 2610.31937
- PDF: https://arxiv.org/pdf/2610.31937
- Categories: cs.AI
- Authors: Yuqing Huang, Jack Ji, Jithin Pradeep, Sagarika Suresh, Jekaterina Novikova
Triage notes (auto)
- paperType:
red-team-vuln - estimatedCaseCount: 2
- triage reason: Automated red-teaming framework demonstrating concrete multi-turn jailbreak successes against six target models with quantified ASR metrics; inherently case-rich for reproducible model-level failures.
0 comments
─────────────────────────────────────────────────────────────────────
// no comments yet