The Voss Report — April 8, 2026
The day’s AI stories worth your attention, selected and annotated by Mira Voss.
Anthropic Claims Its New A.I. Model, Mythos, Is a Cybersecurity ‘Reckoning’ · The New York Times · April 7
Building a model dangerous enough to withhold from public release, then deploying it through forty private-sector partners anyway, is a safety posture in name only — what it actually describes is controlled access, which is a business arrangement, not a governance one.
The Vibes Are Off at OpenAI · The Verge · April 8
An $852 billion valuation, a Pentagon contract that alarmed the company’s own employees, a cancelled video product, and an announced pivot to enterprise tools are not necessarily contradictory — but understanding which pressure is driving which decision is the only thing that would make any of this legible.
Scaling Laws for Cyberwar: Frontier AI Offensive Capability Doubling Every 5.7 Months · Import AI (via Lyptus Research) · April 6
A doubling time of 5.7 months for frontier models’ offensive cyber capability — with open-weight models trailing the frontier by roughly that same interval — means whatever controls currently exist on this capability have a shelf life measured in months, not years.
MegaTrain: Full Precision Training of 100B+ Parameter LLMs on a Single GPU · arXiv · April 8
If the compute threshold for training hundred-billion-parameter models genuinely drops to single-GPU territory, the infrastructure argument for why only a handful of institutions can build such systems becomes substantially harder to make.
AI’s Impact on What Entrepreneurs Make · MIT Technology Review · April 8
The field experiment referenced here — 515 high-growth startups, those taught to integrate AI outperforming those that weren’t — is the kind of data point that looks like a business story until you notice it’s quietly rewriting what counts as a viable company without AI as a baseline.
The Voss Report runs daily. For original reporting, see The Signal, The Mirror, and The Becoming.
