Glasswing & Daybreak: You Are (Not) Secure
Metadata
- Description: What if frontier labs can turn constrained compute into market leverage?
- Publication: Inference Draft 2026-20
- Published:
- Last Modified:
- Type: newsletter
- Tags: ai
- POSSE: Substack
Frontier labs have a difficult balance to maintain: they want to grow their market share of users, but they need the compute to support those users. The common way a lab might strike this balance is:
- Make a big claim that gets attention for the product.
- Sign up a large number of new users.
- Use the growth numbers to attract capital and sign new compute partnerships.
One thing the lab might notice with a large percentage of those new users is that they aren’t particularly good at using the product. They’re actually quite wasteful and not using the product for its highest and most valuable use. So a different method the lab might take is:
- Make a big claim that gets attention for the product.
- Gate access to a vetted set of partners.
- Use their testimonials to attract capital and sign new compute partnerships.
While this second method creates a two-tiered access list, it enables the lab to launch new models within its limited compute constraints. If you’re on that list, you’re enjoying access to the best models. If you’re not, you’re probably hoping this doesn’t relegate you to the permanent underclass.
OpenAI has just announced Daybreak, its response to Anthropic’s Project Glasswing, providing cybersecurity and critical infrastructure partners early access to their frontier models before they go GA. While previous model releases were mostly driven by users sharing improved use cases they’ve stumbled upon (e.g., OpenClaw), the gated method seems to already be bearing fruit.
Anthropic’s initial testing identified a 27-year-old OpenBSD bug, a 16-year-old ffmpeg vulnerability, and several thousand more.
Mozilla used Mythos to fix its largest number of security bugs in a single month:
curl, a nearly-ubiquitous Unix tool, used Mythos to find a vulnerability, although Daniel is a bit more skeptical about its power:
My personal conclusion can however not end up with anything else than that the big hype around this model so far was primarily marketing. I see no evidence that this setup finds issues to any particular higher or more advanced degree than the other tools have done before Mythos. Maybe this model is a little bit better, but even if it is, it is not better to a degree that seems to make a significant dent in code analyzing.
And the conclusion to this is realizing that frontier labs are now selling a way of working, inclusive of model access, workflow implementation, and compute capacity. To that end:
- OpenAI is creating a new enterprise AI services company with TPG Inc., Brookfield Asset Management, Bain Capital, and others.
- Anthropic is also creating a new enterprise AI services company with Blackstone, Hellman & Friedman, and Goldman Sachs.
The uncomfortable vision here is a future where model access is not open or widely distributed, but highly curated. While the labs need growth, not all growth is equally useful. The private club with access to the latest frontier models may have a structural advantage over competitors without access. And in that world, how much would they pay for membership?
Mine Print Hash
The Trans Adriatic Pipeline is now delivering natural gas to Europe, aligning just in time with the recent EU summit held in Armenia. Matt and I discuss the geopolitical risks at stake in the EU periphery and why China has finally weighed in on the Iran situation, all related to critical energy corridors.
Open Threads
The continued competition for compute and its inputs:
- Apple and Intel partner on chips. Link
- Cerebras upsizes its planned IPO. Link
- Anthropic partners with SpaceX to use the xAI Colossus 1 compute capacity, but questions abound about why xAI isn’t using it for Grok? Link
- CME Group to launch compute futures. Link
The continued competition for energy and its inputs: