Re:Qwen4-Max
by jacksonastone on Sunday November 15, 2026 @11:52PM
(#50514178)
Anyone else worried these models have poison pills embedded in the weights? (Past examples: https://cohere.com/blog/concerning-behaviour-popular-open-we...)
[ Parent
| on HN ]
Re:Qwen4-Max
by rohansood15 on Sunday November 15, 2026 @11:55PM
(#50514190)
No technically serious person is buying the Cohere FUD you posted. I'd implore you to dig deeper and come to your own conclusion.
CISOs are buying it. There is no way to prove a model hasn't been poisoned. It's a serious implication that the behaviour can be manipulated at any time. People with heavy compliance obligations are (understandably) nervous because their customers are demanding answers. If I agree with the risk is separate from whether or not the concerns are founded in a technical understanding of the risks.
> There is no way to prove a model hasn't been poisoned.
Which is why the Cohere FUD was so efficient. That's a bullshit claim, but a bullshit claim that cannot be proven wrong can still do massive damage.
So you guard the environment, like you should either way. Any model can and will eventually leak data given the opportunity, intentionally or not.
I'm not too worried when models are popular enough, I assume it will be detected and it can be investigated because it's open, and the reputation of a company doing that would be ruined. Afaik Cohere isn't sure what these urls were made for and it looked more like a way to test the model. Also if you have anything sensitive for that to be a threat you should use a good firewall to detect and block it.
I have more problems when it's a random quantized or finetuned model on hf.