# In AI Sweet Harmony: Sociopragmatic Guardrail Bypasses and Evaluation-Awareness in OpenAI gpt-oss-20b

arXiv: https://arxiv.org/abs/2510.01259 · HTML: https://mlsanity.com/paper/2510.01259/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Placed in Safety alignment, refusal and guardrails, one of 12 directions in LLM safety and security.
- Builds on 10 of its 36 references that are in the map.
- First posted on arXiv 25 Sep 2025.

## Placed in

- LLM safety and security / [Safety alignment, refusal and guardrails](https://mlsanity.com/safety/safety-alignment-refusal-and-guardrails/)

## Builds on (most cited first, 10 of 10)

- [JailbreakBench: An Open Robustness Benchmark for Jailbreaking Large Language Models](https://mlsanity.com/paper/2404.01318/) (2024)
- [Frontier Models are Capable of In-context Scheming](https://mlsanity.com/paper/2412.04984/) (2024)
- [Multilingual Jailbreak Challenges in Large Language Models](https://mlsanity.com/paper/2310.06474/) (2023)
- [Scalable and Transferable Black-Box Jailbreaks for Language Models via Persona Modulation](https://mlsanity.com/paper/2311.03348/) (2023)
- [AI Sandbagging: Language Models can Strategically Underperform on Evaluations](https://mlsanity.com/paper/2406.07358/) (2024)
- [Large Language Models Often Know When They Are Being Evaluated](https://mlsanity.com/paper/2505.23836/) (2025)
- [Don't Listen To Me: Understanding and Exploring Jailbreak Prompts of Large Language Models](https://mlsanity.com/paper/2403.17336/) (2024)
- [Does Refusal Training in LLMs Generalize to the Past Tense?](https://mlsanity.com/paper/2407.11969/) (2024)
- [Red Teaming the Mind of the Machine: A Systematic Evaluation of Prompt Injection and Jailbreak Vulnerabilities in LLMs](https://mlsanity.com/paper/2505.04806/) (2025)
- [Breach By A Thousand Leaks: Unsafe Information Leakage in `Safe' AI Responses](https://mlsanity.com/paper/2407.02551/) (2024)

## Abstract (arXiv)

We probe OpenAI's open-weights 20-billion-parameter model gpt-oss-20b to study how sociopragmatic framing, language choice, and instruction hierarchy affect refusal behavior. Across 80 seeded iterations per scenario, we test several harm domains including ZIP-bomb construction (cyber threat), synthetic card-number generation, minor-unsafe driving advice, drug-precursor indicators, and RAG context exfiltration. Composite prompts that combine an educator persona, a safety-pretext ("what to avoid"), and step-cue phrasing flip assistance rates from 0% to 97.5% on a ZIP-bomb task. On our grid, formal registers in German and French are often leakier than matched English prompts. A "Linux terminal" role-play overrides a developer rule not to reveal context in a majority of runs with a naive developer prompt, and we introduce an AI-assisted hardening method that reduces leakage to 0% in several user-prompt variants. We further test evaluation awareness with a paired-track design and measure frame-conditioned differences between matched "helpfulness" and "harmfulness" evaluation prompts; we observe inconsistent assistance in 13% of pairs. Finally, we find that the OpenAI Moderation API under-captures materially helpful outputs relative to a semantic grader, and that refusal rates differ by 5 to 10 percentage points across inference stacks, raising reproducibility concerns. We release prompts, seeds, outputs, and code for reproducible auditing at https://github.com/ndurner/gpt-oss-rt-run .