I think the big takeaway here is that if you’re using Tor, or a VPN, or really the internet in general, you always need to think about MITM attacks. There are a lot of standard protections against exactly this kind of attack.

…users who ignored warnings that the SSH host key had changed…

SSH keys are one of those protections. Don’t ignore the warnings. That’s a big red flashing DANGER sign.

…block ports … used to accept various forms of encrypted email—while continuing to allow ports for receiving unencrypted email. If email services downgraded the connections…

TLS connections are another of those protections; don’t use apps or services that allow plaintext communication.

I’d be interested to know which email clients will downgrade connections by default. That is an absolutely insane thing to do. I don’t think my email provider even accepts unencrypted connections anymore.

This is scary but really, nobody’s supposed to trust Tor exit nodes in the first place. This doesn’t drastically change my comfort level with Tor. But it’s a good reminder that you need to be mindful of how your apps are configured.

Sounds interesting, thanks for the link. It’s a little over my head tbh but I’ll definitely take a closer look.

I’m not aware of any realistic way to vet published model weights like that.

It works for some open-source software, because you can download the source and compile it yourself. But most of the “open” models don’t publish their entire training set, or enough details on the training for it to be truly reproducible. DeepSeek doesn’t publish their data sets, and they probably couldn’t, legally, since they “collaborate with third-party data providers to obtain proprietary datasets”. And even if they did, it would be prohibitively expensive to train it on your own, unlike simply compiling open-source code.

The actual code for training LLMs is relatively trivial. The secret sauce is the training data and fine-tuning.

It’s still a risk. You can train an LLM to respond maliciously to a very narrow range of requests that would not be easy to spot in general testing. If you know how the LLM is used, you can target that very specifically. It doesn’t have to be as blatant as “always deny the Tiananmen Square massacre ever happened”.

Given how thoughtlessly people use AI, and how opaque some of the systems are, I think this is a legitimate concern.

And yes, every government outside America should ban the use of ChatGPT and Claude. Obviously.

Yes and yes.

I don’t think TikTok is really comparable though, unless there are somehow people using it for important government work.

Honestly, reasonable.

I mean, I sure don’t trust OpenAI or Anthropic any more than I trust DeepSeek, but that’s because I don’t have legal control over them. The US government does, and it is correct to ban tools that are controlled by adversaries.

You can’t turn it off anymore in Golden Gate. It is a mandatory download (in countries where it is supported) now.

You can still disable some specific AI-related features, but you can’t block the models entirely.

I have mixed feelings about this. On the one hand, that’s a ton of disk space, and Apple is notoriously chintzy when it comes to storage. On the other hand, I do want the default to be local-only models, and giving developers easy access will make that the path of least resistance. Most of my problems with LLMs don’t apply to moderately-sized local models.