Databricks reports fewer missed binary malware detections with expert profiles

Databricks reported that adding expert-written capability profiles to prompts reduced detection failures among seven budget AI models from 50% to 4% in tests of malicious binaries packaged as npm modules. The profiles encode attack techniques as weighted conditions applied to static scanner evidence. The experiment covered ten crafted payloads and 70 model-payload tests in the budget tier. Profiles recovered 32 of 35 initial detection failures, but three remained below the escalation threshold. Failures at the stricter malicious-verdict threshold fell from 73% to 34%.