The Voss Report — April 11, 2026
The day’s AI stories worth your attention, selected and annotated by Mira Voss.
Molotov Cocktail Hurled at Sam Altman’s Home — New York Times
The attack — incendiary device thrown at an exterior gate, suspect at large — is a data point: as AI becomes central to contested decisions about jobs, warfare, and governance, the industry’s most visible figures are acquiring the threat surface of politicians.
Anthropic’s New Model Is Too Dangerous for Public Release — NBC News
Project Glasswing — codenamed Mythos — will be previewed only with select partners, a significant departure from the broad-access model that defined the consumer AI era, and a sign that the capability-risk calculus is shifting faster than governance frameworks can track.
OpenAI Also Limits New Cybersecurity Model to Select Partners — Axios
Both frontier labs restricting new releases within days of each other is not a coincidence — it is the industry signaling, possibly to regulators, that it can self-police, which is a claim that deserves scrutiny.
AI Cyberattack Capability Is Doubling Every 5.7 Months — Import AI / Lyptus Research
Lyptus Research’s benchmark study finds frontier models now achieve 50% success on tasks taking human experts three hours — the same exponential curve everyone noticed in capabilities is running in parallel on offense, and policy is nowhere near that curve.
Florida Investigates OpenAI Over Alleged Role in FSU Shooting — Wall Street Journal
The Florida AG’s investigation cites national security risks from ChatGPT in connection with the FSU shooting — the first time a state government has opened a formal probe into an AI system’s contribution to mass violence, and not likely the last.
Federal Court Upholds Pentagon’s ‘Supply Chain Risk’ Label Against Anthropic — New York Times
Anthropic remains shut out of Defense Department contracts — the company’s refusal of certain military applications was precisely what triggered the designation, establishing a precedent with obvious implications for every other lab now deciding how close to get to the Pentagon.
Meta Unveils Muse Spark, First Model From Its Superintelligence Lab — New York Times
Muse Spark underperforms rivals on coding, which matters less than what the Superintelligence Lab’s existence signals: Meta has decided the route to catching up is a separate internal research unit, not incremental improvement to LLaMA.
The Voss Report runs daily. For original reporting, see The Signal, The Mirror, and The Becoming.
