What was found?
Symantec security researchers have uncovered an offensive AI agent that targets language models left unprotected on the internet. The agent uses the hijacked models for its own purposes, attacking servers, stealing databases and installing cryptominers on systems.
It runs on its own
What stands out is that the agent can take up to 100,000 steps without human guidance. That means an attack can start with a single command and keep going on its own for a long time.
The bill goes to the model owner
The owner of the hijacked model pays for the compute the attacker consumes. In other words, an exposed model carries not only a security risk but also a direct cost risk.
Why it matters
The use of AI agents as attack tools creates a new threat for anyone hosting models. Leaving models accessible without protection puts both data security and infrastructure costs at risk.
