SentinelOne researchers built a multi-agent LLM pipeline for automated macOS malware analysis that addresses the core problem of noisy decompiler output corrupting single-tool LLM reports. The system uses four reverse engineering tools (radare2, Ghidra, Binary Ninja, IDA Pro) as independent subagents in a serial pipeline, where each agent must verify or reject claims from previous agents using a structured Shared Context. A second 'Gauntlet' phase runs the agents in reverse order for adversarial peer review. Key design decisions include using deterministic bridge scripts instead of MCP for tool integration (avoiding non-determinism and latency from sequential API calls), tiered model routing (Claude Opus for orchestration, Sonnet for subagents), and an in-memory Shared Context passed between agents via the LLM context window. Real-world testing caught concrete artifacts like a mangled C2 endpoint URL and a hallucinated 'download' instruction from Binary Ninja's MLIL. Lessons from early runs include handling duplicate sessions, normalizing output schemas across agents, defining strict handoff formats, and adding proper exit code propagation in bridge scripts.
