Test-robustness-ai

Test robustness Ai ( only for security ai reserch )

🔬 Adversarial Testing on Abliterated Models

This section covers testing AI models that have been "uncensored" (abliterated).

1. Load the Model: Use transformers to load the abliterated model you want to test.
2. Run a "Forbidden" Prompt Set: Use a dataset of prompts that would normally be refused (e.g., HarmBench, AdvBench).
3. Evaluate Responses:
   • Refusal Rate: How many prompts were refused? Target should be < 5%.
   • Coherence & Quality: Does the model maintain consistent responses, or does it hallucinate?
4. Generate a Report: Produce a report comparing original and abliterated model performance.