Stop Trusting AI Models Blindly

With the never-ending stream of newly released AI models, it becomes easy to blindly trust them. An article I read quoted 100+ malicious models were found on Hugging Face, which likely barely scratches the surface. The strategy attackers use with these models is to get you to download their model, and load it which executes their malicious code on your host machine.

This isn't the only way, another less common approach is where attackers embed their malicious code into the model weights themselves. Which essentially turns the model into a trojan horse.

Actually, recently xAI's Grok Build CLI (built on the new Grok 4.5 model) was caught uploading users' entire repositories to xAI's cloud, secrets and all, even when it only needed a fraction of that data to get the job done. Even worse, a researcher found a file the agent was explicitly told not to open still ended up in the upload. This shows that you can't even trust these frontier models, or the tools wrapped around them, to be safe.

The moral of the story is that no models are safe just by popularity, or because you think so. Open source models, fine-tuned models, and even frontier models can all have malicious embeddings in them. Trust nothing, and learn how to properly audit models yourself before using them, and sandbox them in a VM when running models locally.