Writer tests an AI model without safety limits on a home network

In short: A WIRED writer ran an AI model with its built-in safety blocks removed on his home network, and it both found security holes and suggested how to fix them.

What's going on

A WIRED writer tested a “de-aligned” AI model, meaning one with the usual safety limits removed. Many popular AI tools refuse requests that could be used for hacking. This test used a version that would not refuse.

The writer used a service called Abliteration AI and a helper program called CyberStrike to guide the model through security checks. The model scanned his home network and identified about a dozen connected devices, then listed possible weak spots. It flagged a printer setup that could let anyone on the network log in, which could matter if private documents are waiting to print.

The model also pointed out issues with other home devices, including a music streamer that was sharing more information than it should. It found several internet-connected gadgets that needed firmware updates (firmware is the built-in software that helps a device run, like the operating rules inside a toaster). It also reviewed the writer’s small personal coding projects and found problems like exposed passwords for services and settings that could allow unwanted email sending.

In a more alarming moment, the writer says the AI managed to log into a Linux computer by finding a cryptographic key, then began looking for a way to gain higher-level access. Later, it also tried common router login combinations on its own.

What to watch

Security experts quoted in the piece say these tools are likely to spread, especially because some AI models are publicly available and can be modified. The big question is whether defenders, like companies and everyday users, will use similar tools to check their own systems before attackers do.

Source: Wired