Report: Nvidia in talks to acquire Reflection AI, maker of Beam
FT says a deal could come within weeks, clouding Beam's open-weight drop. Also: an agent sandbox goes GA, a free Cloudflare agent, and on-device PII redaction.
Copy markdown
Nvidia circles Reflection AI as Beam's open weights loom
The FT reports Nvidia is in early talks to acquire Reflection AI or deepen its stake, with a deal reportedly possible within weeks — though talks could still collapse. Nvidia has already put in $800M and Reflection last raised at a $25B pre-money valuation; its 501B-param open-weight coder Beam (23B active) is due to drop weights later this month, and whoever owns Reflection could reshape how — or whether — that release lands.
Microsoft's agent sandbox, MXC, hits general availability
Microsoft Execution Containers are now GA: OS-enforced boundaries that cap which files, networks, and tools an AI agent can touch and block agent-written code from escalating its own permissions. You pick the isolation level — process, Windows session, WSL, or an experimental MicroVM — and learning/permissive modes let you watch behavior before enforcing, but it's launching for commercial customers, so individual-dev access isn't confirmed yet.
Build this weekend: a personal AI agent that runs free on Cloudflare
Talorys is a new MIT-licensed, single-user assistant that deploys to your own Cloudflare account in one command (npx create-talorys@latest) and lives entirely on the free plan — Pages, Workers, and Durable Objects, with Workers AI (glm-4.7-flash) for chat. You get persistent memory, notes/tasks/projects, and recurring reminders on DO alarms; when the daily AI quota runs out chat pauses, but everything else keeps working.
Strip PII in the browser before it reaches any model
Rampart is a 14.7MB ONNX classifier plus a regex/checksum layer that redacts names, cards, SSNs and more entirely client-side — ~3.9ms on WebGPU, ~12.6ms on WASM, 98.4% recall on private terms — swapping sensitive values for placeholders that never leave the device. It's CC BY 4.0, so you can drop it in front of any chat UI as a privacy pre-filter; just know non-Latin names and composed attacks can still slip past.