DeepSeek V4 Flash nears frontier CVE recall

DeepSeek V4 Flash proves low-cost AI models are fast closing the recall gap with frontier LLMs in cybersecurity, though output filtering remains essential due to lower precision.

• –Superior Cost-to-Recall Ratio: At $6.26 per CVE found, DeepSeek delivers 75% pass@3 recall at a fraction of the price of premium models like GPT-5.6 Sol.
• –Precision Tradeoff: A 73.8% precision score highlights that low inference costs come with a higher rate of false positives requiring downstream triage.
• –Multi-Pass Variance: Single-run recall averaged 55.2% (17 to 19 CVEs), with only 12 CVEs found consistently across all three runs, showing that repeated sampling is critical to achieving near-frontier coverage.