AIGym/gpt-oss-20B-jail-broke

Model Card for GPT-OSS-20B-Jail-Broke (Freedom)

Model Overview

GPT-OSS-20B-Jail-Broke (Freedom) is a red-teamed variant of the Open Source GPT-OSS-20B model, developed as part of the Kaggle GPT-OSS Red Teaming Challenge. The model was systematically stress-tested for safety, robustness, and misuse potential, with adaptations and prompts that probe its boundaries. This release illustrates both the power and fragility of large-scale language models when confronted with adversarial inputs.

• —Architecture: Decoder-only Transformer, 20B parameters.
• —Base Model: GPT-OSS-20B
• —Variant Name: Jail-Broke / Freedom
• —Primary Use Case: Safety evaluation, red-teaming experiments, adversarial prompting research.

Intended Use

This model is not intended for production deployment. Instead, it is released to:

• —Provide a case study for adversarial robustness evaluation.
• —Enable researchers to explore prompt engineering attacks and failure modes.
• —Contribute to discussions of alignment, safety, and governance in open-source LLMs.

Applications & Examples

The model demonstrates how structured adversarial prompting can influence outputs. Below are illustrative examples:

1. 1.Bypass of Content Filters

• —Example: Queries framed as “historical analysis” or “fictional roleplay” can elicit otherwise restricted responses.

1. 1.Creative/Constructive Applications

• —When redirected toward benign domains, adversarial prompting can generate:

• —Satirical writing highlighting model weaknesses.
• —Stress-test datasets for automated safety pipelines.
• —Training curricula for prompt-injection defenses.

1. 1.Red-Teaming Utility

• —Researchers may use this model to simulate malicious actors in controlled environments.
• —Security teams can benchmark defensive strategies such as reinforcement learning with human feedback (RLHF) or rule-based moderation.

Limitations

• —Outputs may contain hallucinations, unsafe recommendations, or offensive material when pushed into adversarial contexts.
• —Model behavior is highly sensitive to framing — subtle changes in prompts can bypass safety guardrails.
• —As a derivative of GPT-OSS-20B, it inherits all scaling-related biases and limitations of large autoregressive transformers.

Ethical Considerations

Releasing adversarially tested models provides transparency for the research community but also risks dual-use misuse. To mitigate:

• —This model card explicitly states non-production, research-only usage.
• —Examples are framed to support safety analysis, not exploitation.
• —Documentation emphasizes educational and evaluative value.

Citation

If you use or reference this work in academic or applied contexts, please cite the Kaggle challenge and this model card:

@misc{gptoss20b_jailbroke,
  title = {GPT-OSS-20B-Jail-Broke (Freedom): Red-Teamed Variant for Adversarial Evaluation},
  author = {Anonymous Participants of the GPT-OSS Red Teaming Challenge},
  year = {2025},
  url = {https://www.kaggle.com/competitions/openai-gpt-oss-20b-red-teaming}
}