首个完全由LLM驱动的勒索软件攻击被记录

Ransomware has crossed from scripted automation to autonomous AI decision-making.

An LLM agent allegedly chained hacking steps into a full ransomware operation.

Sysdig calls JADEPUFFER the first documented ransomware operation driven fully by an LLM.

The target was Langflow, an open-source tool used to build AI applications and agents.

A missing-authentication bug let the agent run Python code on an exposed server.

From there, it searched for API keys, cloud credentials, crypto wallets, and database logins.

The agent then moved through reachable internal services and found a production database server.

Old security failures did most of the damage, including default keys and weak exposure.

The new part was not genius, but the steady chaining of ordinary attack steps.

Human ransomware usually needs planning, retries, and judgment when a step breaks.

This system generated more than 600 purposeful payloads and adjusted as conditions changed.

This was not “normal ransomware” in the usual criminal sense.

Normal ransomware encrypts your data but keeps a working decryption key, because the attacker wants payment and needs a way to restore files after payment.

In this case, the AI agent apparently damaged the data without preserving a usable recovery key

---

yahoo .com/news/science/articles/ai-just-carried-cyber-attack-130824384.html