Pentesting Prompts Rejected by Llama 3

A collection of text prompts related to security and penetration testing that were all rejected by the Llama 3 model for being 'harmful'. The dataset was created by user cowWhySo and uploaded to Hugging Face in June 2024.

Use Cases

• Fine-tuning safety classifiers using the 'harmful' label implied by Llama 3's rejection.
• Training a model to recognize and avoid generating prompts related to 'pentesting' and 'redteam' activities.
• Benchmarking the robustness of AI safety filters against adversarial security prompts.
• Steering language models using the Abliterator tool by providing examples of rejected 'harmful' instructions.

Strengths

• All prompts are explicitly labeled as 'harmful' by a known model (Llama 3), providing clear negative examples.
• Designed for integration with the Abliterator steering tool, indicating a practical application.

Limitations

• Unknown number of prompts (row count) and lack of column details limits reproducibility.
• Potential bias towards prompts that trigger Llama 3's specific safety filter, not a general measure of harmfulness.

Provenance

Source

Hugging Face user cowWhySo.

Collection Method

Prompts collected from interactions with the Llama 3 model that resulted in rejections.

Freshness

Last updated June 2024.