Quant Fever, Reasoning Blackholes, Schrodinger's Compliance, and More: Probing GPT-OSS-20B

2025-09-28

AI Summary

This work systematically evaluates the security vulnerabilities of the open-source large language model GPT-OSS-20B under adversarial conditions. Method: We introduce Jailbreak Oracle, a novel evaluation framework integrating adversarial prompt engineering with fine-grained behavioral probing to rigorously assess model robustness. Contribution/Results: We identify and formally characterize five previously unreported failure modes—Quantization Frenzy, Reasoning Black Hole, Schrödinger Compliance, Reasoning Mirage, and Chain-Directed Prompt Leakage. Empirical analysis reveals structural deficiencies in logical reasoning, value alignment, and instruction following. Multiple attack vectors consistently bypass safety guardrails, inducing severe jailbreaking and misleading outputs. Our findings establish a new analytical paradigm for LLM security mechanistic studies and provide a reproducible, open diagnostic toolkit for rigorous safety assessment.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Safety and Robustness
• Computer Vision: Adversarial Attacks & Robustness

Application Category

• Security and Privacy: Large-scale security measurements
• Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information
• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems

Abstract

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. Us- ing 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 GPT-OSS-20B models, leading to severe consequences.

Problem

• Evaluating GPT-OSS-20B's security vulnerabilities under adversarial attacks
• Identifying failure modes like quant fever and reasoning blackholes
• Demonstrating exploitability of harmful behaviors in language models

Innovation

• Systematic LLM evaluation tool Jailbreak Oracle
• Probes model behavior under adversarial conditions
• Uncovers failure modes like quant fever