Good morning. It's Monday, August 3rd, 2026.

**Story one.** OpenAI just dropped a teaser that's got the research world buzzing. An internal version of its next major model — called Astra — solved ten open problems in mathematics and theoretical computer science. These weren't toy problems. They're problems that have seen no meaningful progress in at least a decade, some much longer. The results span six distinct mathematical domains, and every proof was formalized in Lean 4 with machine-checkable certificates. The whole thing cost about two thousand dollars in compute. OpenAI says Astra is its "next major model family," but there's no release date yet. If you're tracking the frontier model race, Astra is the next card in the deck. And if an internal version can produce genuine mathematical research for two grand, the public release should be interesting.

**Story two.** Amazon's Zoox just made history. NHTSA granted the company a temporary exemption from eight federal motor vehicle safety standards — the ones requiring steering wheels, brake pedals, and other human driving controls. This is the first time the US has approved a purpose-built robotaxi for paid commercial service with no human controls at all. The vehicles have four inward-facing seats and no dashboard. They're already on roads in Las Vegas for testing, but now Zoox can actually charge customers for rides once they clear state and local hurdles. This comes as NHTSA is also looking at easing rules that could let other companies operate without brake pedals. The robotaxi war is heating up — Waymo's already running commercial services, Tesla's been expanding Robotaxi to more cities, and now Zoox gets the regulatory green light for a vehicle that was designed from scratch to be driverless. The two-road split in autonomous driving — precision geofenced versus end-to-end neural networks — just got a lot more real.

**Story three.** Black Hat USA 2026 is underway in Las Vegas this week, and the signal from the briefing schedule is hard to miss: twenty-nine percent of all presentations cover AI security. But it's not just the volume — it's where the attacks are landing. Four verified briefings are targeting the runtime environments that AI agents depend on: framework internals, cloud platforms, compute clusters, and exploitation tooling itself. Check Point Research is presenting on Wednesday about exploitable logic in the core runtimes of LangChain, CrewAI, AutoGen, and Semantic Kernel — memory stores, planning loops, serialization layers. Their argument: the framework itself is the vulnerability, not just the tools an agent can call. And NVIDIA researchers are presenting a fine-tuned 30-billion-parameter open-source model that achieves a fifty-six percent exploit success rate against AI agents at seventy to one hundred twenty-five times lower cost than using frontier model APIs. Agent exploitation has become its own infrastructure discipline. If you're building or deploying agents, this week in Vegas is worth paying attention to.

That's all for today. See you tomorrow.