Good morning,

First of all - I got many reactions to last week’s blog post. Many thanks to those people who sent me an encouraging email. I guess we’re doing this.

So I spent yesterday on camera for an hour, talking about Mythos - the AI model that may or may not change everything. The video is below. Watch it if you want the full context. This is just what I'm left with after.

Truth is, I went in thinking I'd play it cool. Ten minutes in I caught myself saying "watershed moment" and "the people who have been asleep should now be wide awake."

So much for cool.

It’s very strange. I had rehearsed some lines to recite but then the thing started and I forgot literally hundreds of people were following along

Mythos wasn't built for cyber security. No security context, no hacker manuals fed in. And it still found a vulnerability in OpenBSD. The operating system that has been bragging "only two remote holes in the default install" for 27 years. 27 years. Gone in a sandbox.

The way they did it bothered me almost more than the result. A thousand parallel agents, $20,000 in tokens burned, and one of them - one - found the bug for $50. It’s a brute force with infinite patience. And that bothered me to be honest because it’s not the smartest model but the one you throw the most money agains.

One of my colleagues on the panel made a point I keep chewing on. Mythos is the product. AI is the risk. We named the webinar after one model, but next month it's something from another lab, then another, then whatever's being trained right now that we don't know about yet. Mythos is just the one that woke us up.

Well, some of us. You , the reader, should have been wide awake for a while now.

What I tried to land on the recording, and what I'll say here too - I'm worried about the asymmetry. Attackers don't have compliance officers. They don't need ethics committees. They don't have to keep a human in the loop. We do. So while someone spins up a thousand agents and goes to bed, our SOC analysts are still triaging tickets one at a time. That gap is going to get violent.

Right around minute 40 I let my inner cynic loose. Something about how it'll take another big breach before boards wake up. The optimist on the panel kept pulling us back to "the basics still hold." He's right. MFA, segmentation, patch management, training. None of that gets less important because Mythos exists. It gets more important. AI doesn't jump air gaps. It just exploits the holes you already had, faster.

But here's the thing I didn't quite say on camera, so I'll say it now.

Most of the corporate audiences I speak to still don't know any of this is happening. They've heard of ChatGPT. They have not heard of Mythos. They are not having board-level conversations about agentic attack chains. They are still arguing about whether to let employees use AI for emails.

That's the gap. Not the technical one. The awareness one. Industry is sprinting. Boardrooms are walking. Regulators are crawling.

Something very bad is going to happen in that gap.

So what about you? When's your next board meeting?

Good morning.

OpenAI Fast-Tracks AI Agent Phone With MediaTek Chips
OpenAI is accelerating its AI-native smartphone to mass production by the first half of 2027, with MediaTek locked in as the sole chip supplier. The device could ship as many as 30 million units in its initial run, signaling serious hardware ambitions beyond software. (Ming-Chi Kuo)

Thiel-Backed Panthalassa Raises $140M for Ocean AI Data Centers
Panthalassa, backed by Peter Thiel, has secured $140 million to build wave-powered floating data centers positioned in international waters. The nodes are designed to use ocean energy to power compute at scale, bypassing land-based infrastructure costs and regulatory constraints. (Business Wire)

Pentagon Expands AI Partnerships, Leaves Anthropic Out
The Department of Defense added eight new AI partners -- SpaceX, OpenAI, Google, Nvidia, Reflection, Microsoft, AWS, and Oracle -- to support classified military operations. Anthropic was notably absent from the list, a significant exclusion given its scale and stated interest in enterprise government work. (Washington Post)

Harvard ER Study: OpenAI o1-Preview Outperforms Doctors on Diagnoses
A study published in Science found that OpenAI's o1-preview model correctly diagnosed patients 67% of the time compared to 55% for emergency room physicians. In one striking case, the model identified a necrotizing fasciitis infection 12 to 24 hours before clinicians caught it. (Science)

White House Moves to Block Anthropic's Mythos Expansion
The White House is pushing back against Anthropic's plan to broaden access to its Mythos model from 50 firms to 120. The opposition adds political friction to Anthropic's enterprise growth strategy at a sensitive moment for AI governance. (Wall Street Journal)

Anthropic Commits $200B to Google Cloud Over Five Years
Anthropic has pledged to spend $200 billion on Google cloud infrastructure over the next five years, a commitment that would represent more than 40% of Google Cloud's current annual revenue. The deal deepens the financial entanglement between the two companies beyond Google's existing equity stake in Anthropic. (The Information)

Musk Admits in Trial That xAI Trained Grok on OpenAI Models
Elon Musk acknowledged under oath during trial questioning that xAI used distilled outputs from OpenAI models to train Grok. The admission came during the ongoing legal battle between Musk and OpenAI, and adds a new dimension to claims about model provenance and training practices. (TechCrunch)

That’s it for us this week. Please like and subscribe 🙂

The answer : Soundgarden - Black Hole Sun