James Wilson @scotchandcode

Enjoy! 🎧 risky.biz/RBFEATURES34/

In this week's Risky Business Feature we covered all this and mode. It's a 45 minute technical deep dive into private LLM inference, followed by 45 minutes delving into the social and governance issues.

Private Inference means the provider can host the model, but is cryptographically locked out of the prompt. Not by policy. Not by “trust us”. By hardware-backed controls and verifiable attestation.

So, I spoke with Tinfoil co-founder @t0nyav to better understand how private LLM inference works. How is this even possible when the LLM transcript is just... plain text?!

During inference, your prompt still has to be processed somewhere. So the sharper question is: can the inference provider see it while the model is running? ... you bet they can!

AI privacy conversations tend to focus on whether your prompts are used for training. That matters, but zero data retention != zero data visibility.

On scanning all the HuggingFace training data.

@osxreverser Would love to chat with you about this for a Risky Business (risky.biz) episode. DMs open!

🎧 Full episode here: risky.biz/RBFEATURES33/ 40 min solo pod delving into the real cost of being forced to switch from open-weight to closed-weight models.

7/ The policy question is not simply whether restrictions deny revenue or reduce misuse. It is also whether they push defenders into retooling while attackers keep moving, and whether compliant US operators get cut out of the open-weight innovation loop.

🧵Let's look at the real cost of having to switch from an open-weight LLM to a close-weight API. It's not an easy switch, especially for cybersecurity work. The trust boundary, privacy, capabilities, availability, and most of all model behaviour matters.

FortiBleed is the antidote to vague “AI-powered attacker” talk. AI was wired into reconnaissance, credential processing, prioritisation, tooling, and workflow. They didn't prompt the agent, the agents prompted them. *That* is AI adoption for cyber. 🎧 New Podcast Episode: risky.biz/RBFEATURES32/

Meanwhile we all exhausted our quota 5 prompts in.

Claude @claudeai

a month ago

We're extending access to Claude Fable 5 on all paid plans through July 12.

6K 9K 81K 22.8M 6K

🎧 risky.biz/RBFEATURES30/ 📺 youtube.com/watch?v=nhS5DT…

Mythos on your desk. Source-local code reviews with frontier-like capabilities. @SecReLabs founder Karsten Nohl and I talked for an hour about this in the latest Risky Business Features episode that's available now on YouTube as well as on your favourite Podcast app. This episode is based on Karsten's recent research and blog post "Beyond Fable: Can a Local LLM Replace Cloud AI for Security Code Reviews". As we got talking though, it became clear to me that this is not just for code reviews. It's equally as useful as a way to construct pen test, patch review, and coding agent harnesses that address cost, privacy, and sovereignty concerns. Use the larger LLM for tasks best suited to it, keep the specific code-level tasks running locally on a smaller LLM. How small? Karsten got great results with a MacBook Air with 32GB of RAM. That's crazy. And, it's also the future with increasingly capable open-weight models.