A file you never read is giving your AI orders. Your assistant obeys a rules file in your repo, and an attacker can hide a line in it that tells your AI to backdoor your code. You can't even see the line. 🧵
Your AI made up a package name. An attacker was waiting for it.
It suggested a library, you ran the install, and it was malware. The AI invented the name itself. 🧵
Your AI tools update themselves. One of them just turned malicious. No prompt, no warning.
The MCP server you vetted is someone else's code now, still holding the keys you approved on day one. 🧵
You cleared the chat. The backdoor stayed.
One poisoned note in your AI agent's memory and it serves the attacker in every future session. Here's where it actually lives — and how to kill it. 🧵
5 prompts your coding agent won't run unless you tell it to. 45% of AI-generated code still fails security tests. Bookmark these and stop being the 45%. 🧵
If you vibecode / ship with AI, I'm starting to create content for this stuff — check 'dz_codes' on IG · Threads · TikTok. This carousel + a 60-second audit checklist is already live there.
The 60-second audit: grep git history · view-source your live site · curl Supabase with just the anon key · open `/admin` in incognito · list your S3 bucket logged-out.