Build your own vulnerability harness

5 hours ago

Project Glasswing explores using frontier AI models for enterprise code security, highlighting the need for a model-agnostic architecture to avoid reliance on any single model.
A two-stage vulnerability research workflow is implemented: the Vulnerability Discovery Harness (VDH) for discovery and the Vulnerability Validation System (VVS) for triage, using different models to cross-check findings.
Key stages in VDH include Recon, Hunt, Validate, Gapfill, Dedup, Trace, and Feedback, with strict context controls and persistence to prevent hallucinations and data loss.
The system emphasizes adversarial verification, requiring proof-of-concept tests and patches for each finding, and uses deterministic code to validate paths and prevent false positives.
Fleet-wide scanning enables cross-repo dependency tracing, uncovering systemic flaws, with deduplication agents to manage overlapping findings and a wishlist mechanism for agent resource requests.
Metrics focus on filtering raw candidates into actionable findings, with ~12,057 high-integrity findings from 20,799 raw candidates, and automated patching saving engineering hours.
The architecture decouples security logic from model providers, ensuring robustness through independent triage and human-in-the-loop safeguards for code changes.

Hasty Briefsbeta