Abstract
Traditional software vulnerability management is currently facing a scalability crisis. Centralized databases like the National Vulnerability Database (NVD) are increasingly unable to keep pace with the rapid emergence of new threats, leading to significant delays and incomplete records. Meanwhile, critical security intelligence has shifted toward informal, unstructured sources such as developer forums, technical blogs, and social media. This paper introduces Evolaris, a self-evolving multi-agent framework designed to automate the discovery, interpretation, and validation of software security threats. By utilizing a recursive discovery engine and a collaborative multi-agent system, Evolaris builds a dynamic knowledge graph that learns and adapts to new threat patterns in real-time. Preliminary results demonstrate the framework's capacity to manage over 260,000 vulnerabilities and 30,000 malicious code implementations, offering a scalable path forward for automated security intelligence.
Introduction
The landscape of software security is shifting faster than human-centric systems can manage. For decades, the industry relied on centralized entities to catalog vulnerabilities. However, the sheer volume of modern software releases and the speed of exploit discovery have rendered these slow, manual processes inadequate.
Today, the most relevant security intelligence often appears first in informal spaces—technical discussions on GitHub, deep-dive security blogs, or underground forums—long before it is assigned a CVE or indexed by the NVD. The challenge lies in the nature of this data: it is fragmented, messy, and difficult to verify at scale. Evolaris addresses this gap by moving beyond simple data collection toward a system that understands, completes, and validates threat intelligence through a self-evolving architecture.
The Evolaris Framework
Evolaris is built upon two foundational pillars designed to transform raw, unstructured data into actionable security intelligence.
1. Intelligence Discovery
The framework employs a recursive discovery engine to navigate the "informal" web. It begins with established, trusted sources such as GitHub Advisory or Snyk and follows reference links to their origins—often original blog posts or forum threads. From there, the system uses search engines to identify similar terminology and related discussions, effectively expanding its reach across the broader web to capture emerging threats that have not yet been formalized.
2. Knowledge Evolution Platform
Once data is discovered, it is processed by a collaborative multi-agent system. Each agent is specialized for a specific stage of the intelligence lifecycle:
- Interpretation Agent: This agent extracts entities and relationships from unstructured text to populate a dynamic knowledge graph.
- Completion Agent: Recognizing that many informal reports are incomplete, this agent analyzes historical threat patterns to "fill in the blanks," such as identifying likely affected source code locations or missing metadata.
- Validation Agent: To ensure accuracy, this agent performs root cause analysis and attempts to generate or migrate Proof-of-Concept (POC) code. By testing the threat in a controlled runtime environment, it provides a "reality check" on the reported vulnerability.
- Risk Detection Agent: Leveraging Large Language Models (LLMs), this agent reasons about the context of code implementations. It distinguishes between truly sensitive operations and benign code that may trigger traditional scanners, thereby refining detection rules over time.
Data Integration and Scale
While Evolaris is presented as a roadmap, its current implementation already manages a massive repository of security data, integrating over 30 authoritative sources (including NVD, OSV, and VulnDB) and 100+ open intelligence sources.
Key metrics of the current system include:
- Vulnerability Management: Over 260,000 unique vulnerabilities tracked.
- Malicious Code Repository: A dataset of over 30,000 malicious code implementations, representing one of the largest open datasets of its kind.
- Unified Presentation: The framework successfully normalizes data from 18+ different major vulnerability sources into a single, interoperable format.
Key Insights: From Static Rules to Logical Reasoning
Evolaris represents a paradigm shift from static security tools to self-evolving intelligence. Most traditional scanners rely on fixed rule sets that quickly become obsolete. In contrast, Evolaris updates its internal meta-model as it encounters new threat patterns, improving its ability to interpret subsequent findings.
Furthermore, the framework emphasizes that context is king. By using LLMs for logical reasoning, the system attempts to understand the intent behind code. This reduces the noise of false positives by determining if a sensitive operation is a legitimate risk within its specific implementation context. Finally, by automating the creation of POCs, Evolaris effectively performs the labor-intensive work of a junior security researcher, closing the loop between discovery and verification.
Limitations and Future Directions
As an evolving roadmap, several technical hurdles remain. Conflict resolution is a primary focus; when multiple sources provide contradictory information about a single bug, the system requires more sophisticated logic to determine the most trustworthy data point. Additionally, the automated compilation and environment setup for thousands of diverse software projects remains a significant engineering challenge.
Future work will focus on tightening the feedback loop between the Risk Detection and Interpretation agents, ensuring that the system's reasoning capabilities continue to sharpen as the global threat landscape evolves.
Summary updated