Purpose-Built DeFi AI Systems
Domain-specific agents vs general LLMs: the 92% vs 34% detection gap on DeFi vulnerabilities, the three pillars of specificity, and the build-vs-buy decision.
The 92% vs 34% gap is a domain-specificity story, not a model-size story. DeFi-native failure modes (oracle manipulation, flash-loan economics, MEV, compositional risk) are absent from general training at reasoning depth. A larger general LLM does not close the gap; domain structure does.
Closing the gap does not require retraining the base model. The three pillars — domain tooling (Foundry/Tenderly/RPCs), encoded attack-pattern knowledge, and execution verification — are external to the base model. Retraining or fine-tuning is rarely justified because the pillars capture most of the 92% at a fraction of the training cost.
The pragmatic architecture is hybrid and per-task. The general LLM handles breadth (triage, explanation, reporting) at lower cost; the domain agent handles deep DeFi detection. Making the hybrid explicit, rather than paying the domain-agent cost on every capability, is the cost-optimization decision.
Demonstrated attacks are weapons, not findings. The domain agent's execution-verified exploits are higher-sensitivity than findings and must be handled as exploit code. Its encoded knowledge base is a steering-attack surface if poisoned. Govern both.