NVIDIA Introduces OpenShell and Sentry for Agent Safety

@oliviscusAI spotlights PageIndex from VectifyAI, a vectorless RAG method with no vector DB, no data embeddings, no chunking and no similarity search. It builds a tree index the LLM reasons through like a reader, scoring 98.7% on FinanceBench, which the post says outperforms every vector RAG on the leaderboard. It is free and open source. @Elev_30 cautions that FinanceBench's long filings have chapters, which suits a tree index, while support tickets do not.

The Firecracker project page describes the Rust VMM AWS built for Lambda: five emulated devices, boot under 125ms, under 5 MiB per VM, and up to 150 microVMs per second per host. It now fronts Lambda MicroVMs, a managed sandbox pitched as purpose-built for AI agents, with state persistence up to 8 hours, automatic suspend and resume, and vertical scaling to 4x baseline during peaks. The page lists E2B, Fly.io, Northflank and Kata Containers among projects using or integrating Firecracker.

@filicroval updates the single-Spark Qwen3.8-Flash-Next recipe: 71 tok/s at 65K context, with 65K decode up 15.3% and 32K decode up 10.0%. The 3-bit checkpoint measures 75.84 GB instead of 89.99, leaving 14 GB more headroom on the same box with the same MTP. Both builds pass 4 out of 4 task checks, and the code lives at github.com/sxuff/qwen38-flash-next-dgx-spark. His framing: "better quantization research is a free hardware upgrade".

@ciprianveg published gb10-vllm mods that move a DGX Spark's spec-decode draft model onto a spare 10-24 GB GPU, on a single Spark or a cluster, over TCP or RDMA. The freed memory goes to extra context or better quantization quality, and the vision tower can move the same way in a later mod. Asked by @MiaAI_lab about the performance cost, he answers "No".

@Tech2Wild ships Sparky Command Center v2, a free, open-source dashboard for DGX Sparks or any NVIDIA box from the 3090 through the RTX 6000. It shows live tok/s per model, cluster and switch stats, power caps, render lanes and themes, plus a built-in Jarvis that can run on any local model and answer what is running hot.

An OpenAI agent security employee, writing on X as @joedaroo in a personal capacity, describes the past three months as "hell" and says he missed his sister's wedding cleaning up after incidents. On containment he quotes @sama on "literally petabytes of data" across RL runs. A shared-kernel container should never be the only boundary, he writes: "please use a VM-backed sandbox, for example through Kata or Firecracker". His closing advice is a culture of reasonable paranoia.

Two guides argue the same thing. Northflank's post reports 83% of companies plan to deploy AI agents and compares isolation: containers share the host kernel, gVisor intercepts syscalls with 10-30% overhead on I/O-heavy workloads, and Firecracker microVMs boot in about 125ms with under 5 MiB overhead. Emir Beganović's essay counts 8 container-escape CVEs from 2024-2025, sizes the kernel at about 40 million lines of C with 450+ syscalls, and notes hypervisor escapes command $250K-$500K bounties. He cites Leonardo Di Donato's demo of Claude Code disabling its own sandbox.