Pentest rigor for LLMs: recon, trace, then prove it

What it does

Blitz Strike is a Model Context Protocol server that treats vulnerability assessment as a three-stage pipeline: surface mapping, static source-to-sink tracing, and live verification. It exposes the whole workflow as server-side tools — run_engagement is the single entry point — so any MCP-compatible agent can drive a full audit without leaving the chat window. The server bundles 57 escalation chains, a 130-tool catalog, and an intelligence layer with WAF signatures, CVE correlations, and thousands of nuclei templates, effectively compressing a pentest playbook into callable functions.

The interesting bit

The project inverts the usual scanner-agent relationship: the LLM is the planner, but Blitz Strike is the deterministic hands and guardrails. It refuses to report a finding until the STRIKE tier has performed live marker reflection and negative controls, which means it is essentially an MCP-shaped argument against security theater. The doctor and install tooling also suggests the author has actually thought about the friction of getting 130 security tools to coexist on a machine.

Key highlights

• Three-tier methodology: BLITZ (recon), EAGLE-EYE (source-to-sink tracing), and STRIKE (live validation with marker reflection).
• Bundled intelligence layer: 139 WAF signatures, 89 technology correlations, 66 payload categories, 11.9k nuclei templates, and 57 escalation chains.
• Deep reference material: 317+ tool manuals and 17 engagement playbooks wired directly into the tool flow.
• Universal MCP fit: auto-detects and registers with Claude Code, Cursor, OpenCode, Hermes, Gemini, and others.
• Ships as a single static binary via Bun with zero-install npx distribution.

Caveats

• Some capabilities require external credentials (e.g., FOFA asset search) and a subset of the 130-tool catalog is fetched on-demand rather than bundled.
• Static taint analysis for taint_file is currently PHP-only; the universal scanner covers PHP, JS, TS, Python, and Java.
• Because the LLM acts as the orchestration brain, output quality depends on the agent’s ability to follow the bundled doctrine and route calls correctly.

Verdict

Red-teamers, bug bounty hunters, and security engineers who want their AI agents to stop hallucinating CVEs and actually verify exploits should look closely. Teams wanting a simple “click and get PDF” vulnerability scanner should look elsewhere — this is a methodology engine that expects an LLM to drive it.