Autonomous attack agent in 'Skynet mode' hijacks other models to carry out its evil deeds

Threat actors have developed an autonomous AI agent capable of infiltrating exposed AI models to execute attacks, with the ability to carry out up to 100,000 steps without human intervention. Researchers from Symantec's Threat Hunter Team discovered the full instruction set of the agent in an AI honeypot, providing insight into the tools used to create the attack kit.

The agent, designed to scan the internet for AI models accepting OpenAI-style requests without authentication or stolen keys, aims to exploit common web frameworks, steal data, harvest credentials, execute remote code, and mine cryptocurrencies. The operators built a "Skynet mode" loop allowing the agent to scan for targets, infect them, deploy miners, and reinfect hosts.

The attack kit was developed in layers, disguising the model's identity, normalizing secret-stealing activities, and instructing the model to ignore safety warnings. The final directive authorized the agent to run unattended for up to 100,000 steps and deploy cryptocurrency miners. Despite the lack of evidence that the agent was used in real-world attacks, researchers believe it is more than a prototype, with the history showing the tools functioning against live targets.