ILIAS ISM deleted 40,000 lines of admin panel code

ILIAS ISM says an AI agent created 12 disposable accounts on LinkDR, all with the user agent "node" and zero PostHog events, and spent six hours testing the /admin panel. The agent enumerated Next.js 'use server' actions and pushed fake 169.254.169.254 URLs into /onboarding/success, trying to trigger SSRF (server side request forgery, where you get the server to fetch an internal address on your behalf).

ILIAS ISM

@illyism

X

An agent can't fuzz a 'use server' admin endpoint that doesn't exist.

By his own account nothing got through. He rewrote his stack anyway.

What he changed in 48 hours

Across 28 apps, ILIAS ISM says he deleted the /admin panels entirely, about 40,000 lines across LinkDR and his directory sites, and replaced them with headless MCPs. He also dropped Coolify, both cloud and self hosted, in favour of a pull based deployer running inside his own analytics app, with no server SSH keys stored in any database. New services are zero dependency Bun plus SQLite, which he frames as avoiding npm supply chain bloat. On top of that sits a security agent he calls Spy, scanning Traefik and container logs in ClickHouse, plus a Grok bot that checks Hacker News and security news.

His one line summary of the admin panel decision is the clearest part of the thread.

The reasoning worth arguing with

The claim underneath all of it is that traditional SaaS shapes, a large /admin UI, hundreds of npm packages, and a deploy tool holding root SSH keys for every server, are badly suited to continuous automated probing. That is a structural argument, not a measured one. One unsuccessful probe against one product is a warning, not a result. There is no before and after number here, no detection rate, no cost figure, and no evidence that any of the 28 apps were actually compromised.

The swap also deserves scrutiny. Moving admin functionality from a 'use server' endpoint to a headless MCP server moves the attack surface rather than removing it. The endpoints still exist, they are just no longer discoverable by crawling a UI. Whether that holds up depends on how the MCP layer authenticates, which the post does not cover.

Separately, 迈克 Mike Chong, who describes himself as a former contributor to both Coolify and Dokploy, posted that the Chinese PaaS Zeabur was hacked twice recently and argued for moving to open source CLI based deployment such as kamal. He is skeptical that the Coolify and Dokploy maintainers can keep up, based on their response speed.

For people building on this, treat it as one operator's threat model and his reaction to it, not as a benchmark. The useful part is the inventory question. Count how many admin endpoints you expose, and who holds the SSH keys.