Rohan Paul (@rohanpaul_ai)T2·68 pts
- Over 60% of token consumption in AI security agents during vulnerability hunting is spent on code understanding and trigger logic reasoning, representing the main cost bottleneck.
- Existing generic cost-saving methods are largely ineffective, with only 24.4% of head-to-head comparisons maintaining success rates at lower costs.
- The new tool AVRI achieves significant cost reduction by returning related code snippets and preserving links between program inputs and unsafe operations, enabling efficient context reuse.