AI's biggest secret, stolen, OpenAI Ultrafast, Grok 4.6, & more

TLDR;

Welcome back! This week, we have a little bit of everything. New releases like OpenAI Ultrafast or Grok 4.6, important market events, new ‘crazy’ AI proposals, and, as always, fascinating research.

Enjoy!

CYBERSECURITY

Researchers steal one of AI’s biggest secrets

I wrote a full article on this topic on Medium if you want to read it for free.

When using model APIs, labs hide the 'intermediate thoughts' of the response, the 'reasoning trace,' to prevent distillation. Preventing distillation is fundamental to staying competitive; otherwise, rivals can distill their models’ capabilities into cheaper models.

But researchers found a cybersecurity flaw in frontier labs' APIs that lets users steal those reasoning traces in full.

Funny thing: it's not an elaborate hacking scheme, but mostly poor cybersecurity practice by the AI labs.

But what happened?

In a nutshell, the encrypted reasoning blobs were accepted across models, sessions, and even users. These signals suggest that labs may have been using the same encryption key, or at least the same shared decryption system, across huge numbers of users.

But what does that even mean?

Well, it means one simple way to get your model’s thoughts decrypted is to take the encrypted data, feed it as part of your input for another model or session, and the server still accepted the encrypted trace as valid, even though it came from another conversation or user, so it decrypted the contents, and you simply have to ask the new model to write, word-for-word, what the encrypted data showed..

As a result, hackers (or, potentially, Chinese labs) could take an encrypted trace, feed it to another model from the same lab, and ask it to reproduce it line by line.

Worse, researchers scraped publicly posted agent logs from GitHub and Hugging Face and decoded more than 300,000 reasoning blocks, recovering API keys, passwords, personal emails, and other sensitive information.

TheWhiteBox’s takeaway:

This is actually pretty hilarious. The companies claiming to have what are basically cybersecurity savants internally, models capable of “hacking the world”, aren’t even capable of correctly securing their own systems.

Incompetence, or blatant misrepresentation of how truly powerful their models are?

INFERENCE

OpenAI’s Ultrafast service is here

With the launch of Ultrafast mode in beta testing, OpenAI now offers its frontier model, GPT-5.6 Sol, at 750 tokens/second, 14 times faster than average responses. To do so, they use Cerebras chips under the hood in a disaggregated inference deployment.

GPUs handle inference prefill (processing the initial input until the first word is generated) plus every text the user provides in their turn. This is a very heavy operation computation-wise, ideal for GPUs.

Cerebras WSEs (Wafer-Scale Engines) take care of inference decoding, the part where words appear on your screen sequentially. This part of the process is not nearly as compute-heavy, and instead requires an enormous amount of data movement, making memory bandwidth the bottleneck

This duality has pushed providers toward heterogeneous implementations that combine GPUs with other inference-specialized hardware, like Cerebras, enabling disaggregated inference deployments.

The secret,…

Enjoyed the introduction? Subscribe or sign in to continue with this edition.