NVIDIA showed a fine-tuned 30B open model matching frontier models at attacking AI agents for 70–125x less.

At Black Hat USA 2026, NVIDIA researchers demonstrated a 56% exploit success rate against AI agents, matching GPT-4o, Claude, and Gemini, at 70 to 125 times lower cost with full local privacy (Straiker). The economics of automated agent exploitation had been implicitly protected by frontier API pricing and provider abuse filters. That protection is gone. Cuts both ways: red-teaming your own agents just got dramatically cheaper, and so did attacking them. If you've been putting off adversarial testing on cost grounds, that excuse expired.