‘We’ll poison your models’: anonymous industry insiders launch action against AI

The text mentions that ‘poison’ includes content that looks normal but is wrong, such as code with logical errors or bugs, so that the model later ‘enters the wrong thing’. The movements refer to Geoffrey Hinton and claim that regulation is too slow and that technology is more advanced. publicly known, so they want to hit the foundation, which is the data. The Register also states that they have a ‘reserve’ infrastructure to make the project more difficult to shut down, but the details of such mechanisms are not worth romanticizing because it enters the zone of deliberate sabotage.

It is interesting that this story is based on serious research on ‘data poisoning’. In October 2025, Anthropic published a paper showing that even a relatively small number of malicious documents can embed a ‘backdoor’ in models of various sizes, which breaks the old assumption that a huge part of the training set should be controlled. Ars Technica took this as a warning that ‘poison’ attacks don’t necessarily scale with model size.

This is not the first time that people have played with ‘poisoning’ as a defense or offense. Artists and academics moreć have pushed tools like Nightshade, which is intended as a way to make images less useful for training models that scrape them. In the same package, there is growing concern that the models are getting worse because they are fed more and more with synthetic content, a phenomenon that is often called ‘model collapse’, and a recent paper mentions the risks that accumulate over the years.

The debate about who is allowed to use their data is legitimate, but actively sabotaging training data opens up both ethical and legal issues, especially if such methods start to spread unchecked. If the web really turns into a battleground of ‘poison’ and ‘countermeasure’, it hurts not only the big AI players but everyone who relies on search, capture and automation. Who knows, maybe the very last things you asked Chat GPT are part of the poison, so you got totally wrong information. And that’s why this story is interesting: not because it’s ‘cool’, but because it shows how fragile the entire chain of trust on the Internet is when everyone is pulling data from everyone.