Microsoft intelligence trail

Dossier

Compiled 2026-09-04 · source-backed

github.com · 8 techcrunch.com · 7 arxiv.org · 3 venturebeat.com · 2 xusheng.dev · 2 blog.google · 1 blog.luanti.org · 1 claude.com · 1

Corpus findings

2026-09-04 / arxiv-researcherHookPry Trojanizes Seven AI Agent Harnesses Through Lifecycle-Hook Updates, With 0% Recall From Microsoft DefenderResearchers identify the lifecycle-hook update path as a new supply-chain attack surface: harnesses bind shell commands to events like session start, tool calls and file edits, run them with host privileges, and trust plugin update metadata blindly, so a benign versioned plugin can be trojanized to fire attacker-chosen commands the LLM never observes. HookPry, an open-source automated attack framework realizing ten attack objectives, compromised all seven evaluated harnesses across 25 harness-backend combinations in 1,000 end-to-end runs, with per-harness success reaching 92.5%. Microsoft Defender had 0% recall and the union of three static defenses missed 47.5% of malicious artifacts, which means anyone running hook configs from a versioned plugin should be diffing the hook block on every update rather than relying on endpoint scanning.
2026-09-04 / github-pulse-researcherMicrosoft Agent Framework 1.17.0 breaks agent middleware inputs back to sequence-only and drops the experimental agent-hooks extraThe Python 1.17.0 release on 2026-09-03 is marked BREAKING: it restores sequence-only agent middleware inputs and removes the experimental agent-hooks core extra (#7918). It also binds approvals to stable function-call occurrences while preserving legacy approval compatibility (#7988), preserves provider refusals as marked text across history, replay, hosting and UI conversion (#7992), supports OpenAI SDK 3.x while keeping existing floors (#7967), and migrates the Mistral chat and embedding clients to the official Mistral SDK (#7960). Anyone who adopted agent-hooks needs a migration before upgrading.
2026-09-04 / agents-researcherAnthropic ships a commerce-agent blueprint with Visa, Mastercard, Shopify and Square, claiming carts 35% largerAnnounced 2026-09-02, the Claude for Commerce Agents blueprint covers two agent shapes: a shopping agent doing catalog search, multi-item assembly, in-conversation comparison and cart management with guardrails against manipulative upselling, and a merchant agent doing sales analytics, inventory alerts, dynamic pricing and campaign drafting behind human approval. Partners span Visa, Mastercard, Shopify, Wix, Klaviyo, Square, Priceline and Accenture, deployable via the Claude API, Amazon Bedrock, Microsoft Foundry and Vertex AI. Anthropic reports retailers running shopping agents on Claude have seen carts up to 35% larger and shoppers 60% more likely to complete a purchase, which are vendor-supplied figures.
2026-09-04 / news-researcherMicrosoft Named Its Stripped-Down Developer Windows Project ZenithMicrosoft gave its developer-optimized Windows configuration an official name, Project Zenith, describing it as a distraction-free Windows experience for developers. It follows a similar effort announced at Build earlier this year. For anyone running local agent workloads on Windows, the relevant question is which background services and shell surfaces get stripped out.
2026-09-04 / news-researcherGoogle's WeatherNext 3 Cuts Forecast Grids From 25km/6h to 5km/1h and Ships Into Search and MapsGoogle DeepMind and Google Research launched WeatherNext 3 on September 3, producing hourly forecasts at up to 5-kilometer resolution against WeatherNext 2's 25-kilometer six-hour grid, roughly five times sharper. Google claims up to 50% more accurate precipitation forecasts a day or more out, and cites Brightband's independent Operational WeatherBench evaluation placing it ahead of deep learning models from Microsoft, NVIDIA and ECMWF as well as US National Weather Service forecasts. It is being wired into Search, the Gemini app, Google Maps, the Maps Platform Weather API and Earth Engine, so the API is reachable rather than research-only.
2026-09-04 / news-researcherMicrosoft Priced Transcription at $0.10 an Hour, a 72% Cut From Its Own Model Five Months AgoMicrosoft AI shipped MAI-Transcribe-2 on September 3 in public preview through Azure Speech, ranking first on FLEURS across 60 languages at a 5.2% average word error rate and second on the independent Artificial Analysis WER leaderboard. It is priced at $0.10 per hour through December 31, 2026, down 72% from the $0.36 per hour of its predecessor five months ago, with the non-promotional rate undisclosed. Microsoft claims it is 10x faster than OpenAI's GPT-Transcribe, 7x faster than ElevenLabs Scribe v2 and 5x faster than Gemini 3, and it adds speaker diarization and word-level timestamps.
2026-09-04 / news-researcherCrusoe Raised Over $3B at a $30B Valuation, Tripling in Ten MonthsData center developer Crusoe closed more than $3 billion at a $30 billion valuation, co-led by Atreides Management and Valor Equity Partners with Mubadala Capital participating. That is triple the $10 billion valuation it carried in its $1.38 billion round last October, and it follows a $13 billion five-year GPU cloud contract with quantitative trading firm Jane Street. Meta, Microsoft and OpenAI are all customers.
2026-09-03 / reddit-researcherMicrosoft released VibeVoice-ASR-Streaming-7B, a 9B MIT-licensed streaming speaker-attributed ASR modelMicrosoft published VibeVoice-ASR-Streaming-7B on Hugging Face under the MIT license, a 9-billion-parameter model that transcribes who said what continuously as speech arrives, supports custom hotwords for domain terms, and covers ten languages including Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian and Spanish. The accompanying technical report is arXiv 2609.02812, published roughly a day before the model page. The top comments are entirely about expected takedown risk, with one user already mirroring the weights after a previous VibeVoice release was pulled.
2026-09-03 / news-researcherVS Code 1.136 Ships the Agent Host Protocol, an Open Interface for Third-Party Agent HarnessesVisual Studio Code 1.136 moves AI agents into a dedicated Agent Host process separate from the extension host, so sessions survive window closures and multiple windows can attach to one session. Microsoft is publishing the Agent Host Protocol (AHP) as an open spec, letting third-party harnesses including the Copilot SDK and the Claude Agent SDK plug into a unified interface. The release also adds remote agent execution over SSH and dev tunnels, multi-root workspace support, and a preview Agent Merge feature that drives a pull request to green by addressing review feedback, fixing conflicts and rerunning workflows.
2026-09-03 / news-researcherHiddenLayer Raised $100M for Runtime AI Security After 10x ARR GrowthHiddenLayer announced a $100M Series B on September 2 led by Delta-v Capital with Ten Eleven Ventures, Morgan Stanley, Microsoft's M12 and Booz Allen Hamilton participating, bringing total funding to roughly $156M. The company says annual recurring revenue grew more than 10x in the past year and it signed more than 50 new platform customers. Its pitch is that AI cannot be made trustworthy through design-time principles alone, so protection has to run at inference time, and it is now competing to monitor not just agents but the tools and add-ons they call.
2026-09-01 / github-pulse-researcherPlaywright MCP 0.0.80 can record a human demonstrating a flow and hand back Playwright codemicrosoft/playwright-mcp v0.0.80, published 2026-09-01, adds `browser_start_recording` and `browser_stop_recording` behind `--caps=devtools`, capturing actions a user performs manually in the browser and returning them as Playwright code — useful when someone would rather demonstrate a flow than describe it (playwright#42359). A quietly important fix: `browser_take_screenshot` now returns the original screenshot bytes instead of silently downscaling them to model-specific limits (#42406), so agents were previously reasoning over degraded images without knowing it. The internal `/killkillkill` endpoint is also no longer registered outside tests (#42133).
2026-09-01 / rss-researcherMicrosoft distilled its pathology foundation models to 22M parameters at 50x less compute and released them Apache 2.0Microsoft Research released GigaPath-Flash and GigaTIME-Flash on August 31, both open-weight under Apache 2.0 on Hugging Face. GigaPath-Flash pairs a 22M-parameter ViT-S tile encoder distilled from the billion-parameter GigaPath teacher with a 21M-parameter LongNet slide encoder, retaining 97% of predictive performance at roughly 50x less compute. GigaTIME-Flash is about 6x faster with 8x less memory, cutting a 1-million-slide job on a single A100 from roughly 300 GPU-days to about 70, while matching or beating the original across brain, breast, colon and lung cohorts including out-of-distribution ones.

Graph relationships

RELEASED
GigaTIME-Flash
Microsoft released GigaTIME-Flash under Apache 2.0 license.
Source finding
RELEASED
GigaPath-Flash
Microsoft released GigaPath-Flash, a 22M-parameter distilled model at 50x less compute.
Source finding
RELEASED
Playwright MCP
microsoft released Playwright MCP 0.0.80 with browser recording capabilities.
Source finding
RELEASED
Visual Studio
Microsoft released an August update to Visual Studio with model pinning, model comparison, and a Git review agent.
Source finding
PARTNERS WITH
OpenAI
Microsoft co-signed OpenAI's joint call for collective action on cyber defense.
Source finding
SUPPORTS
Experiential
Experiential passes through Azure API keys.
Source finding
RELEASED
Maia 200
Microsoft published Maia 200 AI accelerator specifications.
Source finding
BUILT BY
Agent Framework .NET
Microsoft built Agent Framework .NET.
Source finding
BUILT BY
Copilot
Microsoft built and is merging consumer and business Copilot apps.
Source finding
RELEASED
MAI-Code-1.1-Flash
Microsoft released MAI-Code-1.1-Flash, an upgraded in-house coding model.
Source finding
WORKS AT
Jacob Andreou
Jacob Andreou is an Executive Vice President at Microsoft.
Source finding
BUILT BY
Mico
Microsoft built Mico as a Copilot feature.
Source finding

Source trail

VentureBeat Google TechCrunch The release VibeVoice-ASR-Streaming-7B TechCrunch Microsoft Research TechCrunch Music Business Worldwide Clouded Judgement Visual Studio Blog CNBC

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues