Daily briefing

AI & DevOps Briefing — Friday, October 16, 2026

Europe/Bucharest Friday morning digest · AI models, agents, Kubernetes, sandboxes, and notable infra.

AI models

Meta open-weights Muse Spark Open: 816B MoE built for coding and agents

Released Oct 15: 816B total / 42B active parameters, natively multimodal (text, image, audio, video in → text out), 1M-token context. Weights on Hugging Face (BF16 and NVFP4); APIs at Together, Fireworks, Baseten, Modal, Groq, Nebius, and OpenRouter. Runs on two Apple M5 Ultra Mac Studios via MLX, and on SGLang, vLLM, TensorRT-LLM, and llama.cpp. Top story on HN (~710 pts).

Meta AI Research·Meta·Hugging Face·HN

Gemini 4 Argon is broadly available with a 1M unified I/O token limit

Google opened Argon to paid Gemini API / AI Studio users, Antigravity (Ultra), and the Gemini app (Pro/Ultra) on Oct 13. Intro pricing is $2 / $10 per million tokens through Nov 30, then $4 / $20. Positioned for long-horizon science, coding, and agentic work; independent bakeoffs so far put it behind GPT-6 Astra and Opus 5.5 on some coding suites.

Google blog·Google

Krea 3: an image model built for editing at scale

Krea launched Krea 3, aimed at high-volume image editing rather than one-shot generation. Strong HN interest (~280 pts) from designers and product teams evaluating it against Midjourney / Flux-class tools.

Krea·Krea·HN

Goodfire: “Neural networks are clean” — interpretability claim lands hard

Goodfire argues trained networks have more structure than the “messy black box” framing implies, with demos that lit up HN (~580 pts) and fueled a wave of X discussion about why some circuits look surprisingly readable.

Goodfire·Goodfire·HN

Agents

Self Labs launches Agent Vault for credential-safe agent payments

Agents authenticate and pay with mock tokens; real credentials and cards are swapped in only at the moment of use, with spending caps, merchant allowlists, and a kill switch. Beta covered Booking.com, Instacart, and Etsy. Aimed at Claude, Grok, and similar consumer agents that need to checkout without holding secrets.

X news·Self Labs / X

AWS Labs aidlc: adaptive workflow steering rules for coding agents

Show HN of awslabs/aidlc-workflows — adaptive rules that steer AI coding agents through longer tasks instead of one-shot prompts. Useful pattern if you’re building org-level agent playbooks rather than free-form chat.

GitHub·AWS Labs·HN

“CompanyOS” framing: shared rules/memory for multi-agent orgs

Yohei Nakajima and Harrison Chase (and tools like Space / CompanyOS.cc) are pushing the idea of a living company OS where agents and humans share evolving rules, processes, and tool usage with conflict resolution and human oversight — beyond a static RAG “company brain.”

X news·X / builders

Kubernetes

Don’t give agents kubectl: practical K8s sandboxing for LLM-generated code

Kubezilla walks a hardened pattern: dedicated namespace with PSA restricted, least-privilege RBAC (Jobs + logs only), ResourceQuota/LimitRange, default-deny NetworkPolicy, and ValidatingAdmissionPolicy CEL checks. Points to SIG Apps Agent Sandbox (Sandbox CRD, gVisor/Kata) when you need long-lived stateful agent workspaces instead of one-shot Jobs.

Kubezilla·Kubezilla·Kubernetes blog (Agent Sandbox)

Agent Sandbox (SIG Apps) hits v0.4.2 for stateful, isolatable agent pods

Sandbox / SandboxTemplate CRDs, optional WarmPool + SandboxClaim for near-zero cold start, stable DNS identity, scale-to-zero, and runtimeClassName for gVisor or Kata when running untrusted code. Built against Kubernetes 1.35; behavior not guaranteed below 1.32.

AI Briefing·SIG Apps / Kubernetes

Sandboxes

Perplexity open-sources SPACE — Firecracker microVM agent sandboxes

SPACE (Sandboxed Platform for Agentic Code Execution) is the Apache 2.0 Rust platform behind Perplexity Computer: each task gets its own Firecracker microVM, credentials and outbound network stay outside the guest, with pause/resume, snapshots, and an HTTP API. Partners CoreWeave, Akamai, and Crusoe will offer it. Perplexity says 9 frontier models failed to escape in 118 attempts; Kind-based local install is documented. Dominant sandbox story on X overnight.

Perplexity blog·Perplexity·GitHub·SiliconANGLE·X announcement

Harbor + Docker Sandboxes: microVM agent evals locally or in the cloud

Docker’s Oct 14 post wires Harbor so every eval task runs in a Docker Sandbox microVM (own kernel), with network policies that allow deps during setup then lock down during the run. Local free tier + Cloud Sandboxes; stable release targeted Oct 23; Compose-in-sandbox arrives with that release. Install: uv tool install "harbor[sbx]".

Docker blog·Docker

Other notable

Modal hits $1B ARR and raises $750M Series D at $15.75B

Modal (compute, sandboxes, training primitives for AI workloads) announced the round led by Accel; usage reportedly tripled since Series C with 15× YoY growth. A signal that agent/infra sandboxes are becoming a first-class cloud market.

X announcement·Modal

AdFection: targeted ads as a zero-click spyware delivery pipe

Citizen Lab documents how Discover Tech turned ad targeting into a pipeline for mercenary spyware — high-signal security read for anyone running browsers, MDM, or agent browsing sandboxes. HN ~430 pts / ~580 comments.

Citizen Lab·Citizen Lab·HN

RPKI now protects 70% of internet routes

IPinfo reports Resource Public Key Infrastructure coverage crossed 70% of advertised routes — slow, material progress on BGP route origin validation for anyone operating edge or multi-cloud networking.

IPinfo·IPinfo·HN

Libreboot 26.10 “Ominous Olive” adds UEFI and Chromebook support

The libre boot firmware release brings UEFI and broader Chromebook coverage — relevant for hardened / FOSS laptop fleets. HN ~265 pts.

Libreboot·Libreboot·HN

Worth watching