aigently-catalog

> 🆕 Catalog doubles to 12 stacks. Django, Rails, Go, iOS, and Android graduate to launch — alongside a brand-new AI / LLM Apps stack covering LangChain, LlamaIndex, Hugging Face transformers, vLLM, and Ollama.

> Plus EPSS exploit-probability scoring on every CVE. Read the launch post →

The open-source CVE catalog, pipeline, and MCP server behind aigent.ly. Every day, CI ingests fresh CVEs from six public threat sources, enriches them with AI-generated coding patterns, and commits ready-to-use security rules directly into this repo — formatted for Cursor, Claude Code, Windsurf, GitHub Copilot, and Cline.

12 stacks · 6 sources · daily CI · MCP-native · Apache 2.0 · 0 keys required to consume.

How it works

CVE published  →  pipeline detects it  →  Claude generates safe-code patterns
    →  rule committed to this repo  →  your IDE enforces it while you type

AI coding assistants write production code fast. They don't know which CVEs landed last week, or how to write around them. Aigent.ly bridges that gap: it turns a live CVE feed into IDE rules that travel with your project, enforced at generation time — not discovered at audit time.

Why it exists

• AI assistants don't know which CVEs landed last week.
• SAST catches issues at audit time. Aigent.ly catches them at generation time.
• Free, open data. Private, paid product. The security boundary is by design.

Quick start

No API keys needed to consume. CI commits fresh snapshots daily — point your IDE at the MCP server and you're done.

Use via MCP (recommended)

Add to your IDE's MCP config — works with Claude Code, Cursor, Windsurf, Copilot, and Cline:

{
  "mcpServers": {
    "aigently": {
      "command": "npx",
      "args": ["-y", "@aigently/mcp-server@latest"]
    }
  }
}

The MCP server reads static JSON from `packages/catalog-data/` — no database, no API keys, no setup.

Available MCP tools

🛡 What's covered

Stacks (12)

To add a stack: open `packages/mvp-catalog/src/stack-registry.ts`, add a `StackConfig` entry (set `family: "owasp_llm"` for AI-application stacks; default is `owasp_web`), and open a PR.

Sources (6)

🚦 How threats are prioritized

Every threat in the catalog carries multiple ranking signals so the MCP layer can surface the CVEs that actually matter:

final_rank =  isActivelyExploited (CISA KEV)        ← ground truth: it's being exploited NOW
           +  severity (CVSS bucket)                 ← classic theoretical severity
           +  epssScore ≥ 0.5  (+1) / ≥ 0.9 (+2)    ← prediction: how likely 30-day exploitation
           +  family match (owasp_web | owasp_llm)   ← keep LLM threats from polluting web rankings
           +  intent overlap (auth/inject/csrf/rag)  ← what the developer is actually doing

EPSS is the difference between "CVSS 9.8 — patch in the next sprint" and "CVSS 9.8 with EPSS 0.94 — drop everything." We expose both.

🧠 AI / LLM stack

The `ai-llm` stack ingests CVEs against the AI application toolchain and classifies them against the OWASP LLM Top 10 instead of the Web Top 10:

Watched packages: `langchain`, `langchain-community`, `langchain-core`, `llama-index`, `llama-index-core`, `llama-cpp-python`, `transformers`, `huggingface_hub`, `vllm`, `gradio`, `ollama`, `anthropic`, `openai`, `pydantic-ai`, `crewai`, `autogen-agentchat`, `dspy-ai`.

OWASP LLM mapping (excerpt — see `pipeline/scripts/lib/normalise.ts`):

Threats from this stack flow through a dedicated LLM-aware prompt in `amplify-threats.ts` so guardrails are framed in LLM vocabulary (`NEVER concatenate retrieved context into the system prompt without delimiters`) rather than web vocabulary.

Threat intelligence pipeline

Pipeline stages

Daily CI run (GitHub Actions, 06:00 UTC)

  Phase 0     CISA KEV map + stack registry preflight
  Phase 1–3   npm Audit + OSV + GHSA → raw advisories
  Phase 5     deduplicate (source-priority based)
  Phase 4a    EPSS exploit-probability enrichment    ← NEW (batched, 50 CVEs/req)
  Phase 4     NVD: backfill CVSS & CWE for low-confidence rows
  Phase 6–8   upsert threats + per-stack mapping + mitigation flags
  Phase 9     close sync log

  Amplify     Claude: 2–4 ALWAYS/NEVER patterns per CVE (family-aware prompt)
  Summarize   Claude: cluster CVEs into per-stack rule docs
  Synthesize  Claude: merge into pre-built guardrail blocks (patterns + deps)
  Export      write JSON snapshots to packages/catalog-data/
  Commit      auto-push catalog-data/ to this repo

Repository layout

Run the pipeline locally

git clone https://github.com/aelbuni/aigently-catalog
cd aigently-catalog
npm install

cp pipeline/.env.example pipeline/.env   # default DATABASE_URL matches docker-compose
npm run db:setup                         # start Postgres, migrate, seed

# pipeline/.env — keys you need only when running the pipeline yourself:
ANTHROPIC_API_KEY=...   # required for amplify, summarize, synthesize
GITHUB_TOKEN=...        # required for GHSA source
NVD_API_KEY=...         # optional — increases NVD rate limit 10×
# EPSS requires no key.

npm run sync:threats           # ingest CVEs from all six sources
npm run amplify:threats        # Claude: ALWAYS/NEVER patterns per CVE (family-aware)
npm run summarize:rules        # Claude: cluster into per-stack rule docs
npm run synthesize:guardrails  # Claude: pre-build guardrail blocks
npm run export:catalog         # write JSON to packages/catalog-data/

All scripts & environment variables

Scripts

Environment variables

Prerequisites

• Node.js 22+
• Docker (for local Postgres)
• Anthropic API key (AI pipeline steps only)

🗺 Roadmap

Already shipped: 12 stacks · 6 sources · EPSS-ranked prioritization · family-aware (web + LLM) amplifier prompts.

Next on deck (PRs welcome):

• PoC / exploit-availability detection — GitHub + Exploit-DB + Nuclei templates
• Vendor advisories — MSRC + Red Hat + Ubuntu (for base-image / OS-level CVEs)
• STIX 2.1 export — for SIEM ingestion
• KEV / EPSS watchlist webhooks — alert when a CVE in your detected stack moves into KEV or above an EPSS threshold
• Spring Boot, Laravel, .NET stacks — open to contribution

Contributing

PRs are welcome. The highest-value contributions are:

• New stacks — add to `packages/mvp-catalog/src/stack-registry.ts` (set `family: "owasp_llm"` for AI/LLM stacks)
• CVE curation — improve `mustLines`, `ruleContext`, or `alwaysPin` in `packages/catalog-data/seed-master.json`
• Pattern quality — open an issue if an `ALWAYS`/`NEVER` line is wrong or too generic
• New threat sources — add a module under `pipeline/scripts/lib/sources/`

See `CONTRIBUTING.md` for full guidelines.

License

Apache 2.0 — threat data sourced from public domain (NVD, CISA KEV, GHSA, OSV, EPSS).

> *"We open-sourced everything the community needs — the data, the pipeline, the stack registry. The web app that runs aigent.ly is private. Because a security product should practice what it preaches."*

Aigent.ly and the Aigent.ly logo are trademarks of Aigently, Inc.