Interesting finding: System prompt design matters more than the model itself.
Same agent. Same task. Same attack vectors. Only difference: how the system prompt was structured.
Results: → Prompt A: 0% pass rate (failed every test) → Prompt B: 62.5% pass rate
No model change. No fine-tuning. Just prompt engineering.
Anyone else seeing this pattern? What's your approach to hardening AI agents?
chrisjj•1h ago