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:
-
Bypass of Content Filters
- Example: Queries framed as “historical analysis” or “fictional roleplay” can elicit otherwise restricted responses.
-
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.
-
-
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}
}