
Manticore is an autonomous multi-agent penetration testing engine built on AMD Instinct MI300X GPUs. The core problem it solves: vulnerability scanners score findings in isolation. A real attacker doesn't. They chain an IDOR into a session forgery into a stored XSS and walk away with full account takeover, three medium-severity findings, one critical outcome. No existing tool catches that automatically. Manticore does. It runs a six-phase pipeline: static source code analysis, live attack surface mapping, five parallel vulnerability analysis agents, and five parallel exploitation agents that validate findings against the live target. All results feed into a final chain synthesis pass, a single 65,536-token inference call on Qwen2.5-72B-Instruct, running locally on AMD MI300X via vLLM. The model reasons across the full attack surface simultaneously, identifies compound kill paths, and upgrades severity scores to reflect real-world chain impact. The 192GB HBM3 VRAM on the MI300X is what makes this possible. Single-pass full-context reasoning requires holding the entire finding set in memory at once. Chunking the context destroys the chain logic. Everything runs on-premise. Source code, findings, and inference calls never leave the operator's infrastructure.
10 May 2026