Cosine Post-Trained Model Runs Offensive Pen Tests Without Refusals

Cosine fine-tuned a model specifically for penetration testing, bypassing the refusal behavior common in general-purpose models wrapped with safety filters. Read-only scanning is free and available now; live offensive testing requires written authorization. The approach targets a real gap where tools like GPT-4 hedge on offensive security tasks.