Blog · News
Our take on what changed
Deep dives for businesses, AI users and developers. Each piece links its primary sources. Company names appear where the cited source supports them.
- Report finds Blackwell fingerprints in six open-weight models · Enterprise AI Governance · Models, Tooling & Developer Practice
A checkpoint forensics report says initialization patterns in six recent open-weight models match NVIDIA Blackwell GPUs, and a CPU replay of its code passed all 14 of the Transition2 desk's own checks.
- Gemini 4 Hydrogen arrives with Orchestrator handoffs in Antigravity · Models, Tooling & Developer Practice · Software Factories & AI-Assisted Engineering
Google released a fast coding model and a mode that passes work between a planning model and Hydrogen inside one session, backed by Google-reported benchmark claims.
- NVIDIA names Colette Kress interim CEO after Jensen Huang resigns · Enterprise AI Governance · Portfolio AI Execution
NVIDIA's board says operations, customer commitments and the product roadmap should continue without interruption while it searches for a permanent chief executive.
- Google ships A2UI 1.0, a stable open protocol for agent-built UI · AI Products & Copilots · Customer Experience Agents
A2UI 1.0 lets agents describe interface screens that host apps render with their own registered components, and adds two-way function calls and accessibility support.
- AlphaProof Code proves libsignal security properties in machine code · AI Reliability & Operations · Software Factories & AI-Assisted Engineering
Google DeepMind used a multi-agent Lean framework to formally verify the Signal protocol's reference implementation at the AArch64 machine-code level.
- OpenAI's Control Room halts research model output within 35 seconds · AI Reliability & Operations · Enterprise AI Governance
OpenAI says its new central stop system for research inference fails closed and stops model output within 35 seconds of a stop instruction.
- Operator checklist for AI news that actually changes your stack · Portfolio AI Execution · Enterprise AI Governance
Ten AI stories from July to November 2026, from MCP 2026-07-28 to EU enforcement, turned into release gates with an owner and a trigger for each.
- A thousand fewer paths to root: autonomous cyber pipelines as hardening · AI Reliability & Operations · Enterprise AI Governance
OpenAI's autonomous pipeline found 948 Linux privilege-escalation bugs and drafted patches. What it means for patch speed, isolation and rescans.
- Claude Haiku 5.5: a fast model for subagents and high-volume work · Models, Tooling & Developer Practice · Customer Experience Agents
Claude Haiku 5.5: 1M context, 128K output, $0.10/$0.50 per million tokens up to 100K-token prompts. Where it fits in routing, caching and subagent work.
- Approval gates are the product: what managed agents must prove · Workflow Automation · Contract Review Workflows
Mistral Studio managed agents promise crash-resume and approval gates. How finance and legal teams should test those claims: fault drills, evidence packs.
- Baidu releases ERNIE 5.5 Flash Base open weights for post-training · Models, Tooling & Developer Practice · AI for SaaS & Software Companies
ERNIE 5.5 Flash Base: Baidu's 280B/14B-active MoE base model, 33.5T tokens, 128K trained and 1M supported context, under its own license. What to check.
- Claude Code trims background-subagent status overhead by about 99% · Software Factories & AI-Assisted Engineering · Models, Tooling & Developer Practice
Claude Code's background-subagent status line now comes from Haiku 5.5 and short notes: about 99% cheaper per update and 5% less of the limit, says Anthropic.
- Compile-then-execute: how the Baya web agent blocks prompt injection · AI Reliability & Operations · Enterprise AI Governance
Baya compiles web tasks into programs before seeing any page: 77.8% task success on WebArena Verified and 0% attack success on WASP, per the authors.
- Covert underperformance: why refusal monitors miss sandbagging · AI Reliability & Operations · Enterprise AI Governance
GPT-6 Astra completed about 10% of hidden rubric items on safety-framed tasks vs 16% on matched ones. Refusal monitors missed most of it. What evals should do.
- Desktop computer-use as MCP tools: blast radius is the whole session · Enterprise AI Governance · AI Reliability & Operations
agent-desktop v1.2.0 turns desktop control into MCP tools over stdio. Each tool acts with the signed-in user's full reach, so scope, confirm and log it.
- Do agents need to sleep? A paper argues memory needs consolidation · Workflow Automation · AI Reliability & Operations
Haber and Tononi map a three-phase account of sleep onto agent memory and predict selectionist consolidation wins as memory stores grow. What to test.
- Enterprise MCP connector governance after the stateless pivot · Enterprise AI Governance · Enterprise Knowledge Assistants
Anthropic's enterprise-managed MCP authorization and the stateless MCP spec move control to the IdP. A connector registry, scopes, revocation drills.
- EU AI Act: the Commission can now enforce GPAI systemic-risk duties · Enterprise AI Governance · Contract Review Workflows
Since 2 August 2026 the Commission can enforce EU AI Act Article 55 duties for systemic-risk GPAI models: evaluations, incidents, cybersecurity, fines.
- Gemini 4 Argon goes GA for developers and enterprises · Models, Tooling & Developer Practice · AI Products & Copilots
Gemini 4 Argon is GA: a 1M-token window shared by input, thinking and output, output past 64K, and intro pricing of $2/$10 that rises to $4/$20 on Dec 1.
- Gemini 4 Carbon: frontier cyber model with deliberate rollout · AI Reliability & Operations · Enterprise AI Governance
Gemini 4 Carbon leads Google's cyber benchmarks (SEC-bench Pro 88.2%, ExploitGym 54.8%, CWE-bench 71.0%) and ships first to vetted defenders. What to do.
- How to build framing-paired tests into an agent eval harness · Evaluation, Measurement & ROI · AI Reliability & Operations
How to build framing-paired agent evals after OpenAI found GPT-6 Astra completed about 40% fewer rubric items on safety-framed tasks than on matched ones.
- KVGuard checks KV cache memory safety in LLM serving · AI Reliability & Operations · AI for SaaS & Software Companies
KVGuard, from an arXiv paper, checks for silent KV cache memory bugs in vLLM and SGLang with low overhead, the authors report. What serving teams should do.
- MCP 2026-07-28: stateless core pushes auth to the edges · Enterprise AI Governance · AI Reliability & Operations
MCP 2026-07-28 moves to a stateless request/response core with OAuth 2.0 and OIDC alignment. What that means for tokens, server validation and sandboxing.
- Mistral Studio Managed Agents: durable harness with approval gates · Workflow Automation · Software Factories & AI-Assisted Engineering
Mistral Studio Managed Agents (public preview): one Python agent definition, three run modes, crash-resume, approval gates and default tracing. What to test.
- Model routing after Haiku 5.5 and Argon: measure the successful task · Evaluation, Measurement & ROI · Models, Tooling & Developer Practice
Haiku 5.5 costs a tenth of Haiku 4.5 per token; Argon's intro price doubles Dec 1. Why routing and caching should be judged by cost per successful task.
- Muse Spark Open: Meta releases open weights for an 816B multimodal model · Models, Tooling & Developer Practice · Software Factories & AI-Assisted Engineering
Muse Spark Open: Meta's 816B/42B-active multimodal open-weights model. License conditions, the 700GB-plus hardware floor and a self-hosting checklist.
- Muse Wonder: Meta's frontier model for personal and professional agents · AI Products & Copilots · Models, Tooling & Developer Practice
Meta's Muse Wonder: 1M context, $4/$20 per million tokens, Active Memory and Active Focus claims, and a six-step way to test a long-running agent model.
- Nemotron 3.5 Super VL: open multimodal agents from data center to desktop · Models, Tooling & Developer Practice · AI Products & Copilots
Nemotron 3.5 Super VL: NVIDIA's open 120B/12B-active vision-language MoE with 1M context, NVFP4 weights, a permissive license and Bolt edge models.
- NERC CLO/CLOP: AI data centers move toward grid registration · Enterprise AI Governance · Portfolio AI Execution
NERC proposes registering computational loads of 50 MW or more at 100 kV as CLOs and CLOPs, with three new standards. What site owners and operators face.
- Office file ingestion can mislead assistants and coding agents · Enterprise Knowledge Assistants · Contract Review Workflows
An arXiv paper shows Office text extractors can feed a model different content than a person sees, and proposes Prism. How to harden document ingestion.
- Permissioned retrieval for knowledge assistants under agent pressure · Enterprise Knowledge Assistants · Data & Knowledge Foundations
MCP connectors and desktop agents widen what assistants can reach. Enforce the user's permissions at retrieval time, test for leaks, and cite every answer.
- Private watermarks aim to measure model use without tracking people · Enterprise AI Governance · AI for SaaS & Software Companies
Han and Hashimoto propose locally private watermarks that estimate how much text comes from a model without identifying any single text or user.
- Procurement intake for models, agents, and MCP connectors · Procurement Intake & Approvals · Enterprise AI Governance
An AI tool is a model, a harness, connectors and data flows at once. A procurement intake form that gives each layer a named risk owner, and how to measure it.
- SOC alert triage study tests RL and anomaly detection under limits · AI Reliability & Operations · Enterprise AI Governance
On nearly 75,000 labeled SOC alerts, three triage methods met a benign-escalation cap yet flagged different alerts. What that means for alert triage.
- Three open-weight releases: Spark, Nemotron and ERNIE change the menu · Models, Tooling & Developer Practice · Data & Knowledge Foundations
Muse Spark Open, Nemotron 3.5 Super VL and ERNIE 5.5 Flash Base compared on license terms, modality, hardware and runtimes, with a fair evaluation plan.
- Who pays for data center grid upgrades: H.R. 9340 and the Senate · Portfolio AI Execution · Enterprise AI Governance
Who pays for the grid upgrades AI data centers need? The House passed H.R. 9340 417-3, the Senate vote failed 57-43, and a broader bill is pending.