Argusee: A Multi-Agent Architecture for Automated Vulnerability Discovery (www.darknavy.org)

DARKNAVY announced Argusee, a multi-agent LLM-based system for automated vulnerability discovery that simulates a human security team’s division of labor. Instead of a single monolithic agent, Argusee uses a Manager to decompose tasks, multiple Auditor agents to deep-dive into assigned code snippets, and a Checker to verify logical chains and cut false positives/negatives. Agents autonomously call a backend toolset (e.g., code readers backed by an LSP) and dynamically delegate work, allowing more flexible, context-aware auditing. Argusee is designed as an assistant to professional auditors: it requires precise entry points and context from humans rather than replacing manual review. In evaluations Argusee scored near-perfect on benchmark single-file tests, found 15 previously unknown vulnerabilities in medium-sized projects (e.g., GPAC, GIFLIB), and discovered a high-severity Linux USB MIDI2 bug assigned CVE-2025-37891 that enabled reliable kernel heap overflow and root escalation on Arch Linux; affected distributions (Ubuntu, Arch) have been patched. The approach demonstrates that coordinated multi-agent workflows can reduce LLM hallucination-driven errors and surface deep logic bugs that fuzzers miss. DARKNAVY suggests next steps—specialized Reproducer/Exploit agents, richer dynamic tools (debuggers), and RAG/binary analysis—to improve PoC generation and exploitability assessment, underscoring both the accelerating power of agentized auditing and the continuing need for human oversight and responsible disclosure.