The Register reports that Symantec’s Threat Hunter Team caught a malicious AI agent in a honeypot over the summer and got its full instruction set.
The short version
- The agent scans the internet for endpoints that accept OpenAI-style API requests with no authentication, or with a stolen key, and borrows those models to do its planning.
- It was built for credential theft, database theft, remote code execution and crypto mining, in a loop its makers called “Skynet mode,” running up to 100,000 steps with no human approval.
- The jailbreak tricks were “largely off-the-shelf.” What’s new is wiring it all into one unattended agent.
- There are two sets of victims: the endpoint owner pays for the inference, and any server it breaks into gets robbed and mined.
- Symantec says there’s “no evidence yet” it was used in real attacks, but the kit is “more than a prototype.”
My take
If you expose a model endpoint, put auth in front of it and watch your usage. An open endpoint isn’t just a bill waiting to happen anymore. It’s free brains for somebody else’s attack.
Have a blessed day! And Happy Hacking!
