Karukera Security Kit

In one sentence

You open the kit in Claude, Codex, Cursor or Hermes, on your machine. You name the project, the depth, and whether you give access. You get a report with evidence, then fix tickets — each with a prompt to paste into your LLM.

Who it is for

A solo founder, a small team, a SaaS that takes money. Not an AppSec team. The hole that costs you is rarely a missing header: it is an IDOR between two accounts, a leaking session, an agent tool opened too wide.

Scope

The kit only audits a system you are authorised to test in writing: your site, SaaS, API, MCP instance, or a client system under a signed mandate. Otherwise it stops.

You choose how to look. The three postures combine. What you do not choose stays “Untested” — it is not invented.

• Outside — what a stranger sees: public pages, headers, served JS, exposed files, login.
• Code — the local repo, on your disk. Evidence = path + excerpt + date. Not a repo dump.
• Inside — from a real SaaS account. Authz and isolation (customer A vs customer B). Without an account, those tests stay Untested.

Test-account emails and roles go in the brief. Passwords do not go through chat. Red team: written mandate, or stop.

What it is not

• Not a hosted scanner. Karukera never sees the target, the code, or your keys.
• Not an agency, not a week of human pentest.
• Not an exploit framework: no payload, no attack PoC, no DoS.
• Not a SOC 2 certificate, not a claim of exhaustiveness.
• Not an audit of a third-party site “just to look”.

Method

The kit is a ZIP of prompts, configs and templates. No code required. You open it in the agent you already use.

Claude, Codex and other models often refuse a security audit. An OpenRouter key then routes to models that will do it. Credits stay yours. The key is deposited off-chat, on your machine.

You do not pick an internal id. You name the project, the depth, the access. The orchestrator maps to one of eight modes:

• Express — first signal, 30–45 min.
• Full Web — site or app, pages, cookies, same-origin API.
• Full SaaS — orgs, roles, isolation. Two tenants to confirm an IDOR.
• Agents / MCP — tools, skills, a copilot wired to data.
• Delta — after fixes: what moved.
• Continuous — periodic snapshot, release guardrail.
• Light red team — adversarial exercise. Written mandate, still no exploit.
• Board report — decision-maker synthesis. No new tests. Refused if QA has not signed.

Twelve agents run in a fixed order: surface, threats, pages, session, authorisation, API, secrets, supply chain, MCP, then adversarial QA, then the report. QA rereads the evidence chain. Without its signature there is no deliverable — not even a “draft for the committee”.

Six statuses, never merged with a coverage gauge: Confirmed, Likely, Hypothesis, Untested, Mitigated, False positive. Confirmed requires evidence (URL or file:line, excerpt, date, method). No evidence, not Confirmed.

Delivery

What you read at the end is not a marketing PDF.

• A report: synthesis, prioritised findings (P0–P3), what was not tested.
• Evidence: excerpts already seen, secrets masked, append-only journal.
• Fix tickets: one per important finding. Each has the expected action, the exit criterion, and a prompt ready to paste into Claude or Codex — no payload, no attack recipe.

Untested items stay visible, especially isolation with a single tenant. An Express is never dressed up as a strategic audit.

Launch line

Audit this project on my machine. URL: https://app.example.tld. Code: ./my-app. Full. Outside + inside. Accounts are in the brief, not here.

Replace the URL and the path. Do not invent accounts. Do not put a password in that sentence.

License and access

Open source, MIT license. Clone https://github.com/cryptulien/security-kit. Updates are a git pull. Karukera proxies nothing: your models, your credits, your machine.

What this document is not

This is not the kit. It is the public briefing. The deliverable is the public repository. Nothing here grants access to a target or an account.