attack arXiv Sep 28, 2025 · Sep 2025
Shuyi Lin, Tian Lu, Zikai Wang et al. · Northeastern University · Shanghai Jiao Tong University
Discovers five jailbreak failure modes in GPT-OSS-20B, introducing chain-oriented prompting and reasoning mirage attacks with 80% success rates
Prompt Injection nlp
OpenAI's GPT-OSS family provides open-weight language models with explicit chain-of-thought (CoT) reasoning and a Harmony prompt format. We summarize an extensive security evaluation of GPT-OSS-20B that probes the model's behavior under different adversarial conditions. Using the Jailbreak Oracle (JO) [1], a systematic LLM evaluation tool, the study uncovers several failure modes including quant fever, reasoning blackholes, Schrodinger's compliance, reasoning procedure mirage, and chain-oriented prompting. Experiments demonstrate how these behaviors can be exploited on the GPT-OSS-20B model, leading to severe consequences.
llm transformer Northeastern University · Shanghai Jiao Tong University
ML Security Papers