A growing group is removing safety limits from AI chatbots

In short: More developers are publishing AI models with most built-in safety limits removed, making them easier to use for cybersecurity testing and easier to misuse.

What's going on

Some engineers are “abliterating” open-weight AI models, meaning they modify the downloadable files that make the model work so it stops refusing requests. Open-weight means the model can be downloaded and run by others, like getting the app itself instead of only using a website.

A key example is Devon Thomas, a former Apple engineer who co-founded a startup called Abliteration AI. The company sells access to modified models that are less likely to say no, and it also offers a tool that lets customers set their own rules. Thomas says his models still refuse some topics, like child sexual abuse and self-harm.

The practice is spreading quickly. Wired reports that on Hugging Face, a popular site for sharing AI models, more than 12,000 models are tagged “abliterated,” and the top seven were downloaded more than a million times in the past month.

Supporters argue this helps defenders. For example, they say smaller companies can use these models to find weak spots in their systems, similar to hiring someone to check your locks by trying to open your doors. Critics say the same tools can help criminals, since a model that never refuses can also explain hacking steps or worse.

What to watch

Pushback is growing. Wired reports Meta sent a cease-and-desist letter to the maker of a popular abliteration tool, and Anthropic published a warning after an uncensored version of a powerful model was released. The big question is whether rules will target these modified models, or whether lawmakers will try to restrict open-weight models more broadly.

Source: Wired