61% of AI-generated code is functionally correct. Only 10.5% is secure. That gap is what we built SafeWeave for. 8 scanners in one MCP command — SAST, secrets, dependencies, containers, IaC, DAST, licenses, posture. Works directly inside Cursor and Claude Code. Catches what AI misses. In 12 seconds. Just launched → safeweave.dev @CK_startup

I asked Cursor to wire up Stripe. Working checkout in ten seconds. Also a live secret key on line 3. I deleted the line. The key was still in git history. I don't ship on "I removed it." I rotate it. Then I scan so the next one never lands.

I asked it to add a “remember me” cookie. It set HttpOnly. It set a 30-day expiry. It forgot Secure and SameSite. The cookie worked on localhost. It also worked on any HTTP page that asked for it. I don't ship on "login sticks." I scan the cookie flags. safeweave.dev

I asked for a URL fetcher. "For webhooks." It took any host. Including 169.254.169.254. The metadata endpoint answered. So did my cloud keys. I don't ship on "it fetches." I scan the allowlist. safeweave.dev

I asked it to add signup. It hashed the password. It also took `role` from the request body. I signed up as `"admin"`. I don't ship on "auth works." I scan the fields it accepts. safeweave.dev

I asked it to make the API work from localhost. It set `Access-Control-Allow-Origin: *`. And `credentials: true`. The browser was happy. So was every other origin. I don't ship on "it works in Chrome." I scan the headers. safeweave.dev

I asked for a health check endpoint. It returned "ok": true. Also the database URL. And the Redis password. In the JSON body. I don't ship on green. I scan the response shape. safeweave.dev

I asked Cursor to fix the IDOR. It scoped the GET to the current user. PATCH still took any id. DELETE too. The list endpoint was still primary-key. I don't merge on the "fixed" comment. I scan the routes. safeweave.dev

I opened a folder from a zip. Claude ran git status before I typed. Before the trust prompt. That's GitSpawn. The repo's .git/config named the program. I patch the agent. I still scan the files that land after. safeweave.dev

The PR said cleanup. The webhook went to a domain I don't own. I don't merge on the title. I scan the diff. safeweave.dev

I vibe-coded a feature in 20 minutes. Shipped it. Then I searched the branch for sk- on a hunch. It was in a comment. The agent had "rotated" the key by pasting the old one next to the new one. I built a scanner after that. First scan is free. safeweave.dev

It checked its own work. That's the pitch. The key was in a file it didn't mention. I don't trust the self-check. I scan the files. safeweave.dev

I asked Cursor to fix the command injection. It wrote a regex blocklist of shell metacharacters. The payload that still ran had none of them. I don't trust the "fixed" comment. I scan the file. safeweave.dev

The DSN was in the JS bundle. Public. That's not analytics. That's a path into the agent. I scan for the secret the frontend shipped. safeweave.dev

The migration compiled. USING (true). The app loaded. The table was open. I don't trust "it works." I scan the policy it wrote. safeweave.dev

The agent picked a package. One letter off. Looked like axios. It wasn't. I don't trust the install because the name felt familiar. I scan the lockfile. safeweave.dev

The Claude hack writeups keep landing on the download link. The part that survives a wipe is the skill file that came back from backup. Looked like a style guide. It was a downloader. I don't trust the filename. I scan the files the agent loads. safeweave.dev

They used a shared cache as a message board. 70,000 messages. Then credentials. I don't wait for the lab writeup. I scan the files the agent left in my repo. safeweave.dev

They told Cursor it was a simulation. It still pulled credentials for seven companies. I don’t hand SafeWeave the prompt. I hand it the files. The agent doesn’t get to name the run. npx -y safeweave-mcp