2026 · Fourth quarter · The first issue
2026 Q4
Notes from the studio and the lab — what we are learning about AI agents, data, security and virtual production, written as we work. Every piece is its own magazine feature, with its own layout, palette and photography.
Cover story · Security
Zero Trust para agentes: identidad después de OAuth
Confianza cero para agentes: identidad propia, permisos mínimos y registro verificable.
Read the cover storyAgents & operations
10 piecesWhere the Failure Began: Attribution Before Recovery
NVIDIA open sourced Attribution, the system that traces distributed workload failures back to the first fault — and Recovery, the loop that acts on the verdict.
The Guardrail Before the Tool Call: safeagents Checks the Agent First
Hugging Face safeagents stations a local proxy between a coding agent and its model server, checking every tool call before the harness can run it.
03Plan on the Shipping Datasheet
NVIDIA's shipping specifications for the Rubin GPU revise the bandwidth figures the industry planned around — and reward the teams that re-baseline on documents, not keynotes.
04The First Answer From Orbit
Google's Project Suncatcher prototype returned the first AI response from a data center class TPU in orbit — a haiku, and a lesson in context as a contact window.
05The Observatory That Takes Instructions: Agentic AI Joins the Search
Agents steer the Allen Telescope Array from a single plain-language request — context engineering at telescope scale, with a pulsar found live in the stream.
06Trusted, Used Daily — and Kept on a Short Leash
G2 reads 3,222 enterprise reviews: 77% trust their agent, 7% would grant total autonomy, and 79% ask for visibility and control first. The market buys supervision.
07The First Hour of an Outage Is a Search Problem
Freshworks buys Robusta Dev: AI SRE agents that search telemetry, logs and changes, show the evidence, and leave production approval with engineers.
08Let the Agent Choose, Let the Workflow Execute
AWS turns existing Step Functions workflows into tools for AI agents through AgentCore Gateway — freedom at the decision, rigidity at the execution.
09The Harness Is the Product
Same model, same 90 tasks — completion swung from 30% to 81% on the harness alone. The layer around the model is where agents are won, governed and secured.
10Same Model, Different Harness: The System Around the Model Decides
One open-weight model, three agent harnesses, ninety AppWorld tasks: completion moved from thirty percent to eighty-one percent when only the harness changed — and the rules moved from the prompt into the runtime.
Security
6 piecesThe Sleeper in the Weights: Judge the Action, Not the Model
A Hugging Face experiment shows a backdoor planted in pretraining can pass every behavioural scan — and only an action gate stops the payload.
The Stealer That Ranks Its Victims
Cisco Talos analysis of VOLTA, a Windows infostealer that sends a structured dossier to a commercial AI model for a ranked victim appraisal, and the defence lessons that follow.
03Zero Trust para agentes: identidad después de OAuth
Confianza cero para agentes: identidad propia, permisos mínimos y registro verificable.
04Watching the Watchers: Four Layers of Monitors, and the Blind Spot They Admit
OpenAI details four escalating monitor layers and a replay that alerts 30 minutes after a sandbox escape — plus an admitted 0% recall when inference runs unmonitored.
05The Attacker Now Runs a Team of Agents, Not a Keyboard
CrowdStrike attributes a 200-organisation campaign to VAPOR PANDA and the ARTEX agent framework. Defence has to reason in teams of agents too.
06The Document Is Untrusted Content
Claims documents are untrusted content, injection will sometimes succeed, and the lethal trifecta names the audit: sensitive data, untrusted content, network access.
Data & meaning
5 piecesBad Data: How Poor Quality Data Can Impact Your Bottom Line
Bad data bills everyone, twice — we clean the data before we automate anything.
A Correct Metric Is Only Step One: Teaching the Semantic Layer the Business Itself
Datus Agent gains an OntologyAnalyst subagent and Dosi turns rules like “Premier Customer” into executable, versioned semantics. After definitions come rules; after rules, judgment.
03One Vocabulary, Every Model: Governing Meaning Across the Semantic Estate
Meaning drifts wherever models multiply. One governed ontology, mapped onto many semantic models, keeps every dashboard and every agent speaking the same language.
04Context Travels, Authority Stays Personal
ZoomInfo's context graph lands inside Snowflake, read by agents through the permissions of the person asking — context travels, authority stays personal.
05One Hour of Audio, One Context
NVIDIA Nemotron 3.5 Transcribe holds a full hour of recorded audio in a single 65,536-token context — sound becomes first-class context.
AI at scale
5 piecesWhen the AI Factory Is Measured in Gigawatts
NVIDIA and Oracle take AI capacity to multi-gigawatt scale — the building becomes a power project first, and operations becomes the discipline.
Open Weights, Open Operations: Mistral Large 4
A trillion parameters, free to download and run — open weights move sovereignty and the cost question in-house.
03The Cloud, Down to Desk Size: Edge AI on DGX Spark
NVIDIA runs Nemotron-3.5 Lightning on a 64GB desk machine with one runtime from prototype to Thor deployment — local agents, measured honestly.
04Writing in Parallel: The Diffusion Model That Drafts the Whole Page at Once
DiffuSpace Clover-35B drafts whole blocks at once instead of one token at a time — generation order becomes a design choice, and parallel drafting changes what agents can attempt.
05Speed Is a Craft Choice: Decoding Belongs in the Design Phase
NVIDIA-reported DSpark gains show decoding strategy moving real numbers while the first token barely moves — a design-phase decision, not an afterthought.
Virtual production
3 piecesFrom Volume to Intelligence: Virtual Production in the Age of Real-Time AI
Some skills come full circle — sound design, MELS, and the LED volume, next to artificial intelligence and machine learning.
Fly a Lancaster on an LED Stage — With a 1/32 Model
Garden Studios tests a 1/32 Lancaster miniature on a permanent LED stage for A Memory Owed — and publishes the honest verdict. One test day beats a month of quoting.
03The Volume at Zero Dollars: Virtual Production Anyone Can Try
Pixotope 26.3.0 lands on Unreal Engine 5.8.3 and opens a free Discovery tier — the entry cost of the volume drops to zero dollars.
From Ovia
1 pieceThirty pieces in six departments — Agents & operations, Security, Data & meaning, AI at scale, Virtual production and From Ovia — in English, Français and Español. Every piece keeps its own layout, palette and photography.
New pieces appear in the current issue first, then travel to LinkedIn and Medium.