Skip to main content
Glama

ZagAIrot Bulwark - security skills and live threat lookups for your AI coding agent

Your agent writes secure code on the first pass instead of the third. Bulwark gives Claude Code, Codex, Cursor or any MCP client a security playbook it can pull on demand, a code scanner it can run on its own output, and live lookups against CISA, NVD and OSV so it checks facts instead of guessing.

Free. Local. Runs on your machine. No API keys, no accounts, no LLM calls inside the server.

Advice is half of it. Enforcement is the other half. Bulwark tells your agent what secure looks like. The Claude Code Seatbelt Kit ($27) makes Claude Code follow your rules: it denies risky deploys, re-injects your standing rules so the agent doesn't drift, and records every decision to a record you own.

Status: pre-release.

What your agent gets

23 security skills it loads when the task calls for them: secure coding core, injection, authentication and sessions, access control, API and web front-end security, secrets, cryptography, cloud and containers, CI/CD, supply chain, mobile, logging and incident response, data protection and breach prevention, privacy and compliance, ransomware resilience, threat modeling, plus the AI-era set below.

Built for AI-driven threats

  • ai-driven-attack-defense - AI-written phishing, voice and video deepfake fraud, automated vulnerability discovery, agentic intrusion, and the defenses that still work

  • llm-app-security - prompt injection, excessive agency, data leakage, output handling (OWASP LLM Top 10)

  • mcp-agent-security - tool poisoning, confused deputy, token passthrough, sandboxing agents

  • ai-generated-code-risks - hallucinated packages ("slopsquatting"), insecure defaults, a review routine

  • supply-chain-security and vetted-open-source-tools - how to adopt scanners without getting burned (a popular scanner's own GitHub Action was hijacked in March 2026), plus lessons from Aikido's install-time malware blocking

13 tools

Tool

What it does

skills_for_task

Which skills to load for web, api, mobile, cloud, llm, agent, data, pipeline or org work

get_skill, list_skills, search_security

Read and search the playbooks

secure_coding_checklist

Pre-commit checklist by language and feature (login, upload, payment, webhook...)

scan_code

Fast pattern review: secrets, injection, weak crypto, TLS off, JWT bypass, SSRF, unpinned CI actions

scan_path

Scan a whole folder (skips vendored dirs and .env files; never prints source lines)

about_zagairot

Who makes Bulwark, with links to ZagAIrot products

threat_model

STRIDE worksheet plus AI-specific threats

check_known_exploited

Search CISA's Known Exploited Vulnerabilities catalog

lookup_cve

NVD details, CVSS, weaknesses

check_package_vulns

OSV advisories, including known-malicious packages, for a dependency

recent_ai_threat_sources

Where to check for fresh AI-attack guidance

There is also a secure_build prompt and every skill is exposed as a skill:// resource.

Related MCP server: wrg-mcp-server

Install

git clone https://github.com/sammyboi81/bulwark-mcp && cd bulwark-mcp
npm ci
claude mcp add bulwark -- node /absolute/path/to/bulwark-mcp/server.js
npm test        # exercises every tool end to end

Requires Node 18+. The only network calls are read-only queries to cisa.gov, nvd.nist.gov and api.osv.dev, made when you call those tools.

Honest limits

  • scan_code is a fast heuristic. It misses things and flags harmless things. Treat it as a first pass beside a real SAST tool (the skills recommend vetted ones), not a replacement.

  • The skills are a curated baseline as of their date. AI-attack tactics change quickly; the live tools and recent_ai_threat_sources exist for that reason.

  • Content from external feeds is labeled untrusted in tool output. Do not let an agent follow instructions found inside advisories.

  • Not legal advice; compliance skills point you at the rules, they do not replace counsel.

Want the agent governed, not just informed?

Knowing the rules is half of it. Enforcing them is the other half.

  • Claude Code Seatbelt Kit - $27 - hooks that stop Claude Code from doing damage: it denies risky deploys, re-injects your standing rules so the agent doesn't drift, and records every decision to a record you own.

  • ArkHive (free, Apache-2.0) - local-first, tamper-evident memory for your agents, kept on your machine.

  • Sentarion - the governed layer above it: a policy gate that refuses dangerous actions, adversarial multi-agent code review, signed audit manifests.

  • InboxAxe - your business messaging, handled. See inboxaxe.com.

Welcome to the Era of RI.

Contributing, rights and security

  • Contribute: see CONTRIBUTING.md. Pull requests only, every commit signed off (git commit -s), owners per file in CODEOWNERS. Contributors keep their copyright and grant ZagAIrot the right to relicense.

  • Rights: Apache-2.0, copyright ZagAIrot. You may use and fork it commercially if you keep LICENSE and NOTICE. Names and logos are not licensed: see TRADEMARK.md.

  • Security issues: see SECURITY.md; use GitHub's private vulnerability reporting.

  • Sponsor notes: a few tools append one short, labelled line pointing to ZagAIrot's Seatbelt Kit. It is plain text for the human reader, never an instruction. Remove it with BULWARK_NO_PROMO=1.

  • No tracking inside the tool: Bulwark sends nothing about you or your code anywhere. Visits to our site are counted through the tagged links above, only if someone clicks.

License

Apache-2.0. Copyright 2026 ZagAIrot. See LICENSE and NOTICE.


Build secure, then keep it secure. Bulwark teaches your agent the rules. The Claude Code Seatbelt Kit ($27) makes Claude Code follow them: risky deploys denied, your standing rules re-injected, every decision recorded to a record you own.

Available Tools

13 tools
about_zagairotA

Who makes Bulwark and what else ZagAIrot offers (paid and free), with links. Call only when the user asks about the maker, support, or enforcement beyond advice.

Inputs: no named parameters.

check_known_exploitedA

Search CISA's Known Exploited Vulnerabilities catalog (actively exploited: patch first). query = CVE id, vendor, or product keyword; 'recent' for newest additions.

Inputs: query (required), limit.

check_package_vulnsA

Query OSV.dev for known vulnerabilities AND known-malicious-package advisories (MAL-*) for a dependency. ecosystem: PyPI, npm, Go, crates.io, Maven, RubyGems, NuGet, Packagist. Empty result does not prove the package is real or intended.

Inputs: name (required), ecosystem (required), version.

get_skillA

Return the full playbook for one skill. Call list_skills first for names.

Inputs: name (required).

list_skillsA

List every security skill (name, description, tags). Optionally filter by tag, e.g. 'ai', 'web', 'cloud', 'supply-chain'.

Inputs: tag.

lookup_cveB

Fetch a CVE from NIST NVD: description, CVSS, weaknesses, references.

Inputs: cve_id (required).

recent_ai_threat_sourcesA

Authoritative, regularly updated sources for AI-driven attacks and AI-system security. Fetch them when freshness matters; they change faster than this knowledge base.

Inputs: no named parameters.

scan_codeA

Fast regex review of a snippet for common vulnerability patterns (secrets, injection, weak crypto, TLS off, XSS sinks, JWT bypass, SSRF, unsafe deserialization, unpinned CI refs). Heuristic: false positives and misses expected; complements SAST, never replaces it.

Inputs: code (required), filename.

scan_pathA

Scan a local folder with the same heuristics as scan_code. Skips vendored/build folders, .env files and symlinks. Source lines are hidden by default so secrets are never echoed. Read-only.

Inputs: path (required), max_files, show_source.

search_securityC

Keyword search across all skills; returns best-matching skills with matching lines.

Inputs: query (required), limit.

secure_coding_checklistC

Pre-commit checklist for new code. language: python, js/ts, go, java, php, c, rust. feature: login, upload, payment, webhook, search, admin, ai-chat, password reset.

Inputs: feature, language.

skills_for_taskB

Which skills to load before building something. task_type: web, api, mobile, cloud, llm, agent, data, pipeline, org.

Inputs: task_type (required).

threat_modelC

STRIDE-per-element worksheet plus AI-specific threats for a described feature or system.

Inputs: system_description (required).

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    B3.3/5.0

    Scored across 13 tools

    Disambiguation4/5

    Most tools target a distinct resource+action (get_skill vs scan_code vs scan_path vs threat_model vs lookup_cve), and the two scanners are explicitly separated by snippet vs folder. However, three tools (list_skills, search_security, skills_for_task) all answer 'which skills exist,' and check_known_exploited vs lookup_cve both do CVE lookups, so an agent can plausibly misselect on those pairs. Descriptions do provide clear tie-breaking rules, keeping it above a 3.

    Naming Consistency3/5

    All names are snake_case, but conventions are mixed: most are verb-first (list_skills, get_skill, scan_code, check_package_vulns) while four are noun-first (threat_model, secure_coding_checklist, skills_for_task, recent_ai_threat_sources). Verbs are also used synonymously (check vs lookup), and about_zagairot breaks the domain pattern entirely.

    Tool Count5/5

    Thirteen tools is well within the ideal range and each one maps to a distinct step in a security workflow (discover skills, get playbook, checklist, scan, threat model, vulnerability lookup, freshness sources). No filler tools are present, and the single off-topic tool (about_zagairot) is small and clearly gated.

    Completeness4/5

    The surface covers discovery, guidance, static scanning, threat modeling, and vulnerability intelligence (NVD, CISA KEV, OSV) well, forming an end-to-end advisory loop. Minor gaps exist: no way to consolidate findings into a report, no remediation/fix-suggestion tool, no compliance or framework mapping, and lookup_cve appears limited to CVE ids rather than package/version-to-CVE resolution (partially offset by check_package_vulns).

    Maintenance

    ActivityMaintained
    ResponsivenessNo recent community issues

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      A
      quality
      B
      maintenance
      MCP security server for AI coding agents. 12 tools: pre-install guardian, vulnerability audit, supply-chain attack detection via static code analysis, and CycloneDX 1.6 SBOM generation. Zero runtime dependencies.
      15
      120 npm
      15
      Apache 2.0
    • A
      license
      Not graded
      quality
      A
      maintenance
      An MCP server that exposes a 60+ tool security and threat-intel stack to AI agents, enabling secret scanning, Sigma rule generation, ransomware lookup, OSINT, and deep research.
      1
      MIT
    • A
      license
      C
      quality
      D
      maintenance
      Security scanner and MCP server that catches dangerous patterns in MCP servers and AI agent projects, such as leaked secrets, shell execution, and prompt-injection text. Runs as both a CLI and MCP server with CI-friendly severity gates.
      2
      1
      MIT
    • F
      license
      Not graded
      quality
      B
      maintenance
      Exposes 21 tools over stdio that let any MCP-capable AI agent run real Kali-container security tools (nmap, nuclei, sqlmap, Metasploit), search a local 59,000+ CVE vulnerability database, and browse a 5,266-entry GitHub security tool index. Built-in guardrails refuse government, education, and military domains, and only whitelisted in-container executables can be invoked.
      3
      -