.png)
Lineaje AI Labs used our Frontier Defense™ harness to assess hundreds of software projects using multiple frontier AI models and then generated verified patches that were compatible with those applications. We define “novel vulnerabilities” as new weaknesses in code that frontier models detect that do not map to known CVEs.
These results demonstrate that code-level vulnerabilities can now be eliminated autonomously with limited human supervision.
Every project contained multiple novel vulnerabilities:
Across hundreds of software projects, Frontier Defense discovered previously unknown vulnerabilities in 100% of projects evaluated, averaging 5.67 novel vulnerabilities per project.
About half of novel vulnerabilities are exploitable:
48% of enumerated novel vulnerabilities are exploitable.
Understanding exploitability requires full application code assessment, including its dependencies and runtime behavior.
Reachability determination is insufficient and unnecessary when you can directly determine exploitability: Frontier Models generate exploits for non-reachable and for reachable novel vulnerabilities. Reachability analysis underestimates real-world exploits proven by frontier models.
Autonomous Fix can eliminate all exploitable vulnerabilities, not just prioritized ones, while reducing both security and developer toil:
Frontier Models harness generated verified, compatible patches 76% of the time. 24% needed developer involvement.
The harness could separate these out and drive elimination in both situations. Overall, Autonomous Fix reduces developer toil by more than 90% for novel vulnerabilities.
Frontier Model Assess and Fix harnesses can be overlaid on top of existing infrastructure to eliminate both traditional and novel vulnerabilities at scale.
Learn how Frontier Defense powers Continuous Vulnerability Elimination Factory (CVEF).
