AI Security Knowledge Base — 2,052 OWASP Check Cards for AI Coding Agents

What it is

secure-kb (the AI Security Knowledge Base) is a retrieval layer built for AI coding agents. It merges 67 tracked sources into 2,052 atomic security check cards, normalises them, and exposes them through a fast gate, a platform router, and a zero-dependency SQLite FTS5 index.

It is not a scanner, and it does not replace a professional audit. What it does is tell an agent what to verify for a given codebase — then let it report pass / fail / n-a with evidence. Every card traces back to a real file, anchor and commit.

Give an agent a codebase — get the 15–25 checks that actually matter, in seconds.

The problem it solves

Ask a coding agent to "review this repo for security bugs" and you usually get generic advice: escape your inputs, use parameterised queries, keep dependencies updated. Correct, but unfalsifiable and untethered from the code in front of it.

The knowledge base flips that around. Instead of a vague prompt, the agent detects the stack, pulls the checks that apply, and works through them one by one — the way a human AppSec reviewer would, but in seconds. Card text is strictly defensive guidance; it never contains exploit payloads.

What's inside

Cards come from official OWASP material and a handful of high-quality community baselines, each with its own licence recorded.

Cards use source-identity IDs such as SKB-ASVS-V3-4-5 or SKB-WSTG-INFO-01, and merged cards keep every historical id in aliases[] so nothing ever stops resolving.

The five-step workflow

It ships as a CLI, an MCP server (9 tools), and an Agent Skills bundle, so it drops into Claude Code, Codex, OpenCode and any MCP-capable harness.

1. Detect the stack — secure-kb detect /path/to/project returns ranked platforms.
2. Profile it — profile web --gate fast.
3. Gate — the top checks by priority, domain-balanced, inside a small token budget.
4. Drill down — get <id>, or search by CWE / ASVS / MASVS reference.
5. Report — pass / fail / n-a / unknown, each with evidence. No "pass" without evidence.

Zero required dependencies — standard-library Python only. The index builds itself on first use.

Verifiable attestation

An agent's security claim is normally unfalsifiable prose. secure-kb can instead emit a Verifiable Security Attestation — a deterministic, tamper-evident record of which checks ran, against which KB build, on which exact code fingerprint — that anyone can verify offline.

Change the checks, use a different KB build, or edit the target, and verification fails. It is also available over MCP as attest and verify_attestation.

Use it

Clone it, run detect, and you have an answer in under a minute. The project is currently a public beta (v0.9.x): the build is deterministic and reproducible, but the retrieval targets are not yet met, so it is deliberately not v1.0.0.