Dev tools

Blitz Strike is an MCP pentest server that live-verifies every finding before it reports it, and it has 638 stars in a week

September 19, 2026 at 6:30 PM PT

GitHub page for the Blitz Strike repository

Image: GitHub

Why it mattersAgent-driven security scans that report unverified findings waste a triage team's day, so a server that treats each hit as a hypothesis and tests it before reporting is directly about that cost.

Blitz Strike, released on 12 September, is a TypeScript Model Context Protocol server that packages a three-tier penetration-testing methodology and holds back every finding until a live HTTP test has confirmed it. The repository was 638 stars 7 days after it went public, from an author with a broader kali-pentest project the tool draws its 317 tool manuals from.

Recon, then trace, then live verify

The server exposes tools in three phases. BLITZ enumerates the attack surface: unauthenticated entry points, dangerous sinks, and authentication boundaries. EAGLE-EYE runs inter-procedural taint analysis in PHP, JavaScript, TypeScript, Python and Java, and traces whether a sink is reachable, unauthenticated and exploitable, rather than merely present in the code. STRIKE runs the live check: strike_verify(url, method, data, headers, marker, timeout) sends the request, reads for the marker in the reflection, and requires the chain's negative_control to stay inert before it will mark the finding confirmed.

The author's own note on the method: "Every finding is a HYPOTHESIS until strike_verify reflects your marker AND the chain's negative_control stays inert." One run_engagement call executes the full path server-side from any MCP client, whether Claude Code, Cursor, OpenCode, Codex or Gemini.

139 WAF signatures, 57 chains, 11,900 nuclei templates

The intelligence layer is data-driven and comes from named MIT-licensed sources: 139 WAF signatures, 89 tech-to-vulnerability correlations, 53 CVE correlations, 103 port-to-service mappings, an attack-vector taxonomy of 34 categories covering 588 vectors, 66 exploit payload categories from PayloadsAllTheThings, and 11,900 nuclei detection templates. On top of that sit 57 escalation chains such as ssrf_cloud_metadata, jwt_alg_confusion_forgery and subdomain_takeover, each carrying an ordered set of steps with success criteria, invariant checks and a negative control.

A 130-tool catalogue covers 15 security categories: 23 recon tools (subfinder, amass, httpx), 13 exploitation tools (sqlmap, hydra, hashcat), 12 reverse-engineering tools (ghidra, radare2, angr), 8 Active Directory tools (netexec, impacket, bloodhound) and more. Each entry has its base command, key flags, per-platform install command, install check and homepage. 317 indexed manuals across 100 percent of the catalogue are wired into the flow, so run_engagement attaches the relevant manual per matched chain.

Detection 1.0, false positives 0, on a self-selected corpus

Blitz Strike carries a labelled corpus of vulnerable and safe fixtures across four languages and a run_enterprise_benchmark tool. The author reports a detection rate of 1.0, a false-positive rate of 0 and precision of 1.0 on that corpus, and a coverage matrix of 56 out of 60 language-by-sink-class pairs, with the four missing pairs legitimately absent from the language. A bun run test runs 183 regression checks on finding lifecycle, CVSS math, taint tracing and benchmark invariants on every push.

A team already running agent-driven security work has been reading unverified findings and paying for the triage. A methodology that refuses to report until it has reflected a marker and inertly failed the negative control is a specific answer to that cost, and the working is a benchmark file rather than a claim. Blitz Strike is one week old, so the benchmark corpus is small and self-selected, and the false-positive number will move once a team runs it against real production traffic.

Source

Source: Blitz Strike

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.

More from AI News

Keenable SELECT is a research agent that runs live web search as DuckDB SQL, with semantic operators inside the query

Keenable has published SELECT, an MCP server that lets a research agent search and process the live web from inside a DuckDB SELECT query, using semantic operators that call an LLM per row only when a plain SQL WHERE has narrowed the set.

Source: Hacker NewsDev tools

pgbot exposes Postgres diagnostics to AI agents through a role-based read-only MCP server

pgbot is a new Apache-licensed Go binary that gives AI agents a read-only Postgres diagnostic tool through MCP, and the repository has gained 870 stars in about three weeks.

Source: GitHubDev tools

OKF Agent Memory gives coding agents a searchable memory with no embedding API

OKF Agent Memory stores agent memory as plain Markdown files in your repository and searches them with BM25 in under 300 microseconds, so there is no vector database and no embedding API to pay for.

Source: Hacker NewsOpen source