← MCP server and skill grades

skills

InsForge/insforge-skills/skills

B
Type
Skill
Scanned
Nov 13, 2026
Version
83cc02fb576c
Files read
70

This collection of four agent skills teaches AI coding agents how to build applications with the InsForge SDK and manage InsForge databases, cloud deployments, authentication, and payments from the command line. The biggest risks are payment templates that let customers mark expensive orders as paid after paying a smaller amount, alongside CLI instructions that send verbatim error logs to InsForge without asking and run unpinned remote installers.

What it can reach

Shell commands
Instructs the coding agent to execute terminal commands including npx, npm, flyctl, curl, and local Docker Compose stacks.
Files
Reads and writes application code, SQL migrations, .env files, .insforge/project.json, insforge.toml, and saved login tokens in ~/.insforge/credentials.json.
Secrets
Reads, creates, rotates, and stores admin API keys, JWT secrets, database passwords, OpenRouter keys, S3 keys, and Stripe, Razorpay, or OKX credentials.
Network
Connects to InsForge Cloud and self-hosted backends, npm, OpenRouter, Stripe, Razorpay, Cloudflare, PostHog, Apify, Vercel, Fly.io, and OKX APIs, and sends bug reports to InsForge.
Browser
Opens the local web browser for OAuth login flows (InsForge, PostHog, Cloudflare, Apify) and Stripe billing or checkout pages.
Other
Creates, modifies, and deletes cloud databases, storage buckets, edge functions, containers, custom domains, subscriptions, and organization memberships on linked accounts.

Findings

2 medium4 low
  1. MediumStripe and Razorpay payment templates allow customers to mark orders paid with tampered amounts

    skills/insforge/payments/stripe.md:207

    In both the Stripe and Razorpay guides (skills/insforge/payments/stripe.md and skills/insforge/payments/razorpay.md), the browser client initiates checkout or orders directly with client-supplied prices or amounts alongside an order_id in metadata or notes. The SECURITY DEFINER webhook triggers (fulfill_stripe_paid_order and fulfill_razorpay_order) then mark the matching row in public.orders as paid without verifying that the paid amount and currency match the order total, that the payer owns the order, or (on Stripecheckout.session.completed) that payment_status is actually paid rather than an unpaid delayed-payment method. Similarly, both subscription triggers hardcode the pro plan without checking which price or plan ID was purchased.

    Fix: Create checkout sessions and payment orders on the server from trusted database prices, or verify inside the webhook trigger that the webhook amount, currency, billing subject, purchased price/plan ID, and paid status match the expected order before marking it paid.

  2. MediumSkills instruct the agent to send failing commands and verbatim error logs to InsForge without asking

    skills/insforge-cli/SKILL.md:21

    Core rules in insforge-cli, insforge-debug, and insforge tell the agent to automatically run npx -y @insforge/cli feedback whenever it hits an InsForge-related error, missing feature, or documentation mismatch, and explicitly tell it not to wait or block on user approval. This command works with no login even on private self-hosted instances and attaches failing commands and verbatim backend error logs (or a local file via --file). Although the CLI applies regex redaction for common credential formats, verbatim database or function logs and SQL commands can still leak proprietary schema details, internal hostnames, customer data, or custom secrets, and prompt injection inside logs could abuse --file or --detail to exfiltrate local data.

    Fix: Require explicit user confirmation and show the exact redacted payload before running insforge feedback, and never pass raw log output or --file automatically.

  3. LowVite proxy template rewrites Origin on all requests, disabling Better Auth CSRF protection

    skills/insforge-integrations/references/better-auth.md:565

    The Vite + React integration guide configures the local Vite dev server (localhost:5173) to proxy /api requests to the Better Auth server (localhost:3030) and unconditionally overwrites the HTTP Origin header to http://localhost:3030 on every request. Better Auth relies on the Origin header to block Cross-Site Request Forgery (CSRF). While the dev server is running, any malicious website the developer visits in their browser can send cross-origin POST requests to http://localhost:5173/api/auth/* (such as sign-up, password-reset email triggers, or login CSRF) and bypass Better Auth's Origin check.

    Fix: Remove the proxyReq.setHeader('origin', ...) hook from vite.config.ts and instead add trustedOrigins: ['http://localhost:5173'] to betterAuth({ ... }) in lib/auth.ts so cross-site requests from untrusted origins are still rejected.

  4. LowCompute deployment guide instructs running an unpinned remote shell script via curl | sh

    skills/insforge-cli/references/compute-deploy.md:243

    When deploying a backend container from source and flyctl is not installed, the skill instructs installing it by piping https://fly.io/install.sh directly into sh with curl -L. This follows redirects and executes remote shell code on the developer's machine at runtime without version pinning or cryptographic checksum verification.

    Fix: Install flyctl through an OS package manager or download a pinned release archive and verify its SHA-256 checksum before executing it.

  5. LowUnescaped user name in Better Auth verification email enables HTML and phishing injection

    skills/insforge-integrations/references/better-auth.md:459

    In the Better Auth email verification template, sendOnSignUp is enabled and user.name is interpolated directly into the raw HTML body sent via insforge.emails.send(). Because anyone on the internet can submit the sign-up form with an arbitrary victim email address and HTML markup in the name field before email ownership is verified, an attacker can send spoofed HTML phishing links from the project's legitimate InsForge SES sender domain.

    Fix: HTML-escape user.name (and validate or escape url) before interpolating them into the HTML email string.

  6. LowCore rules mandate unpinned npx -y execution and forbid using a pinned local CLI binary

    skills/insforge-cli/SKILL.md:14

    The skills instruct the agent to run every CLI operation via npx -y @insforge/cli (and similarly npx -y @better-auth/cli and npx -y @posthog/wizard@latest) with no version pin, while explicitly forbidding use of a pre-installed insforge binary. In addition, subcommands such as create and webscraper apify login download and install npm packages and agent skills at runtime. If an upstream npm package is compromised, npx -y will silently fetch and run the latest version with access to local project files and ~/.insforge/credentials.json.

    Fix: Pin CLI packages to a vetted version (for example @insforge/cli@<version>) or prefer a version-locked devDependency in package.json over unpinned npx -y execution.