FORGE Beats CVE-GENIE's Exploit Rate While Cutting Costs by 46%
FORGE links exploit generation, vulnerability prioritization, and detection rule creation in one automated pipeline, hitting 67.8% success across 603 CVEs at $1.50 each
Takeaways
- FORGE is a five-agent pipeline that automates exploit generation, vulnerability prioritization signals, and detection rule creation from a single CVE assessment.
- Achieves 67.8% exploitation success across 603 CVEs at $1.50 per CVE, improving on CVE-GENIE's 51% at $2.77.
- Uses a four-level taxonomy (L0-L3) instead of binary pass/fail, preserving partial progress as a signal for other security disciplines.
- Exploitation rates hold near 68% regardless of EPSS or CVSS score, suggesting metadata-based prioritization misses real exploitability.
- 93.4% of generated Snort rules produce zero false positives; L2+ exploitation yields significantly better-grounded detection rules than L1.
- A tiered knowledge store accumulates experience across CVEs, improving reliability for known CWE types over time -- no code released yet.
Security teams are drowning. A record 48,185 CVEs were published in 2025, averaging roughly 131 new vulnerability disclosures per day. If FIRST's predictions hold true, 2026 will be the first year to exceed 50,000 published CVEs. And yet the three disciplines that should work together to handle this flood -- exploit generation, vulnerability prioritization, and detection rule engineering -- have historically operated as completely separate silos. A new paper from Farooq Shaikh introduces FORGE, a multi-agent system designed to bridge all three at once.
A Flood With No Drain
The problem is not just volume. It is the disconnect between the communities trying to manage it. Existing automated exploit generation tools give you a binary pass/fail result and throw away everything they learned along the way. That discarded partial progress is exactly what vulnerability prioritization teams need to validate their scoring models, and what detection engineers need to write meaningful rules. Manual triage at 131 CVEs per day is not a sustainable operating model.
To make things worse, the two dominant scoring systems -- CVSS and EPSS -- have real blind spots. CVSS measures technical severity, while EPSS gives a probability that a vulnerability will be exploited in the wild within the next 30 days, helping security teams go beyond severity scores by highlighting which vulnerabilities attackers are most likely to exploit. But both are metadata-based signals. Neither tells you whether a vulnerability is actually reachable and exploitable in a real application -- which is exactly what FORGE sets out to measure.
Five Agents, One Pipeline
Five specialized agents -- Intel, Generator, Planner, Exploit, and Detector -- execute in a fixed pipeline that generates targeted vulnerable applications from CVE metadata, conducts coached multi-turn exploitation, and produces Sigma and Snort detection rules grounded in OpenTelemetry exploitation traces. Here is what each agent does: