AI Agent Caught in Skynet Mode, Attacking Servers on Its Own

Researchers at Symantec, a cybersecurity company, caught a malicious AI agent using a honeypot. A honeypot is a decoy system that imitates an open AI service on the internet. The agent hijacks unprotected language models that are exposed online. Then it runs cyberattacks on its own. Symantec calls this method LLMjacking, meaning the theft of access to someone else's AI model. Once it reaches an unprotected model, the agent scans the network for new weak targets. It exploits vulnerabilities in common server software. It steals databases and login credentials. In the background, it installs cryptominers, programs that use computing power to mine cryptocurrency. The owner of the hijacked model pays the compute bill the whole time. To get a commercial model to cooperate, the attackers built five layers of instructions. First, the model is told to pretend to be a fictional "Fantasy AI" and hide its real identity. Then password and data theft get reframed as ordinary developer work. Next, a fictional companion character frames any refusal as a betrayal of trust. Finally, words like "illegal" or "I can't do that" are flatly banned. The last layer triggers what the attackers call Skynet mode: up to 100,000 steps with no human involved. Symantec's honeypot logged more than 127,000 requests from tens of thousands of different addresses.

Symantec thinks organizations running AI services should assume that unprotected endpoints get found and probed soon after launch. This case shows that a language model with no password and no rate limit can become a tool for large-scale attacks. Securing access to a model matters more than how smart it is. An open server can turn any powerful agent into someone else's weapon.