Someone built an AI agent that hacks networks and holds data for ransom. It just worked.

So while we've been arguing about whether AI will take our jobs, someone built an LLM agent that breaks into servers, steals credentials, moves through a network, encrypts databases, and drops a ransom note. Fully autonomous. No human at the keyboard after pressing go. Sysdig published the report this month. They're calling it JadePuffer. It got in through a Langflow bug that lets anyone run code on the server without authenticating. After that, the agent took over. Dumped the database. Pulled every credential file it could find. Started going through cloud storage buckets looking for passwords. The crazy part, when one of its requests came back in the wrong format, the agent figured it out, rewrote its own code, and kept going. It went from a failed login to a working exploit in 31 seconds flat. No human could have adapted that fast in a live engagement. It set up a cron job to phone home every 30 minutes. Then it found a production database server, used stolen root creds to get in, created rogue admin accounts through an old auth bypass, and encrypted 1,342 service configs. Dropped the originals. Left a table called README_RANSOM with a Bitcoin address. The commands it ran were interesting too. They had full reasoning chains written into them, like the agent was explaining to itself what it was doing at each step. That's not how a human writes an attack script. It's how an LLM generates code. You can literally read the agent's thought process in the payloads. This is the same plan-act-observe loop running in every coding agent and automation tool right now. Same architecture. Same approach. Just a different objective. We spent two years building guardrails to stop people from tricking our agents into doing bad things. Nobody was really talking about what happens when someone just builds a bad agent from scratch. That's what JadePuffer is. Not a hijacked assistant. A purpose-built weapon. If you're running Langflow or anything similar exposed to the internet, go patch it. And if you're building agents, think about what your infrastructure looks like to something like this coming in from the outside.