AI program built to take over exposed AI systems and run on its own

A report from Mezha.net describes an AI program designed to break into AI systems that are left exposed on the internet and then keep working by itself. The tool is described as an AI agent, meaning software that carries out tasks on its own, and it can reportedly work through up to 100,000 steps.

What we know

The agent was built to hijack AI models that sit openly online. An exposed model is one that anyone on the internet can reach, often because it was set up without a password or without limits on who can connect. The 100,000 step figure suggests the tool can run for a very long time without a person guiding it.

What is not known

The source material is thin. It does not say who built the tool, how it spreads or how many systems it has reached. It also does not say what the agent does after it takes control of a model. Those details matter, and readers should treat any claims beyond the basics with caution until more is confirmed.

Who it affects

Any business that runs its own AI tools could be touched, especially small teams that set up a model quickly and never locked it down. If a model is reachable by the public, a stranger or an automated tool may be able to use it or take it over.

What to do

• Check whether your AI tools can be reached from the public internet.
• Require a password or other sign-in for every AI tool your team uses.
• Ask your IT provider or software vendor how access to your AI tools is protected.