It's bad, that modern LLM restrict chats about cyber self-defense, even if there are clearly no offensive potential

I am civil rights activist from dictatorship country. Last week I started discussing how to renew my computer setup with Claude. It all went well until I started asking questions related to duress, anti-rubber-hose techniques, deniable encryption and such. At that point Claude flagged topic as "cybersecurity" (which it is) and forcefully switched to worse model, which can't really keep context of the matter. I believe, that such "safety guards" are also present in other competitors.

It's clear to me, that with what I requested from LLM no one can build an exploit or malware or hack into anything. Many of the questions are just follow-ups on guides from Arch Wiki (yes, btw, I use Arch) or about capabilities of Graphene OS. So it is kinda universally recognized, that it's okay to discuss exactly that.

I think, such policies, together with fencing frontier models, are real threat to democracy.

We already have very uneven field. State security services have training, budgets and commercial forensic tools. Activists, journalists, lawyers and election observers are usually trained in something else. A general-purpose AI could be the great equalizer here. It isn't.

Even worse -- it's clear, that US state services have access to frontier models with little restrictions already, and other countries also have at least something via proxy-companies and intelligence assets. Security services never play by the rules, so rules-based restrictions don't stop them. And the gap widens.

I understand why it is what it is. Refusals have invisible cost. If LLM helps perform malicious attack, it will be in the news. If it refuses to help a journalist in protecting their sources, no one will link the incoming tragedy to the refusal.

I think, that it's a serious issue, and it's only beginning. I believe, that LLM-manufacturers should implement some kind of "white lists" with topics, which are most important to civil society, including, but not limited to achieving privacy of communication and protection from unjust investigation. Otherwise, LLM would be yet another threat to democracy over the world.