AI Weekly Brief — week ending 2026-09-12

This Week in AI

DeepSeek ships V4.1-Flash and rattles chip markets. The new open-weight model cuts memory and API costs, and coverage says it beats Claude Opus 5 and GPT-5.6 Sol on several tests. Samsung and SK Hynix fell over 3% intraday on fears the design needs less HBM. TechRepublic, Bloomberg
Anthropic accuses seven China-based labs of distilling Claude. Anthropic says the labs, including DeepSeek, Moonshot AI, and Alibaba, routed at least 35 million user requests to Claude through fake accounts to train rival models. The Hacker News, Business Insider
Moonshot AI targets $2 billion in annual revenue. The Kimi maker is weighing a Hong Kong IPO at a reported $50 billion valuation, a sign that open-weight labs are building real commercial engines. TechCrunch, Bloomberg
New models reach enterprise platforms within days. DeepSeek V4.1 Flash is already live on Databricks Unity Gateway and on Baseten's model APIs with a 1M-token context window. Model governance processes now have days, not months, to evaluate new releases. Databricks Release Notes, Unite.AI
GitHub Copilot code review gets more autonomous. Copilot now resolves its own review comments when a later commit addresses them, and enterprise usage metrics now cover the VS Code Agents window. Teams measuring AI adoption get better data. GitHub Changelog, GitHub Changelog
FTC rescinds its 2021 health app breach policy statement. The Commission withdrew guidance on breaches by health apps and connected devices. Health plans with member-facing apps should watch how enforcement posture shifts under the Health Breach Notification Rule. FTC
Microsoft 365 Copilot needs firewall updates and had an access incident. Teams and Copilot are moving to new network addresses, and Microsoft investigated Copilot access issues under incident CP1470554. Both matter for any org that depends on Copilot daily. Computerworld, CyberSecurityNews

Vendor Dossiers

Alibaba (Qwen)

Anthropic names Alibaba, a China-based lab, among those it accuses of training on Claude output to close the gap with US frontier models. No new Qwen release appeared in this week's intake. International Business Times

Anthropic (Claude)

Anthropic says seven China-based AI labs ran industrial-scale distillation attacks on Claude, using fake accounts to route at least 35 million user requests through its models over the summer. The disclosure names DeepSeek, Moonshot, and Alibaba. The Hacker News, Business Insider, Chosunbiz
Reactions split: Y Combinator CEO Garry Tan said he would not take action against model distillation if it happened to him. UA.NEWS

DeepSeek

Released: DeepSeek V4.1-Flash, an open-weight model that cuts memory and API costs, with coverage citing test results above Claude Opus 5 and GPT-5.6 Sol. DeepSeek, a China-based lab, shipped it ahead of a planned Shanghai listing. TechRepublic, GIGAZINE, technology.org
The model's sharp cut in HBM and SSD usage hit memory-maker stocks: Samsung and SK Hynix fell over 3% intraday while Micron and SanDisk held up. The Register argues it sets a template for capable models that run lean. Yahoo Finance, The Register
Pricing pressure is spreading: coverage frames V4.1-Flash as a fresh blow to OpenAI and Z.AI in a price war now going global. The Business Times, The Rundown AI

Google DeepMind (Gemini, Gemma)

No Gemini or Gemma items appeared in this week's intake.

Mistral

Announced: Mistral and Cloudera partner on sovereign enterprise AI, bringing Mistral models to enterprise data in place. Mistral is France-based, which anchors its sovereignty pitch for regulated industries. TNGlobal, AiThority
Announced: Loft Orbital and Marlan Space plan a $1 billion AI satellite fleet with Mistral. The Next Web
Mistral's enterprise pitch centers on control over deployment and data rather than raw model capability, per a TechTarget profile. TechTarget

Moonshot AI (Kimi)

Moonshot AI targets a $2 billion annual revenue run rate on the back of its open-weight K3 model, and is weighing dual Hong Kong-Shanghai listings at a reported $50 billion valuation. The China-based lab's trajectory shows open-weight strategies producing real revenue. TechCrunch, Bloomberg, mezha.net
Moonshot is one of the labs Anthropic accuses of siphoning Claude output through fake accounts, a provenance question hanging over its models. Chosunbiz

OpenAI (GPT, Codex)

OpenAI published two GPT-6 Astra customer stories: Perplexity uses Astra for end-to-end work on communications, software changes, and production monitoring, and Cognition uses it to help Devin test its own code. Both stress less human review per unit of output. OpenAI, OpenAI
An engineering post details how OpenAI scaled its Habitat storage platform to serve 1 billion ChatGPT users at 22 million requests per second. OpenAI

Zhipu AI (GLM)

Zhipu, a China-based lab, ended its promotional pricing just as DeepSeek cut prices again, deepening the global price war among Chinese model providers. finance.biggo.com
Head-to-head coverage compares GLM 5.3 Flash against DeepSeek V4.1 Flash on benchmarks, pricing, and speed. Intelligent Living

GitHub Copilot

GA: Copilot usage metrics now report activity in the dedicated VS Code Agents window, with daily active users and per-user session counts at enterprise and organization level. Useful for anyone tracking Copilot adoption across teams. GitHub Changelog
GA: Copilot code review now auto-resolves its own comments once a later commit addresses them, writes commit messages for applied suggestions, and uses more shell tools plus an agent ensemble to check its work. Open comments now reflect only feedback that still needs attention. GitHub Changelog

Microsoft 365 Copilot

Teams and Copilot are moving to new network addresses; IT teams must update firewall rules to avoid breakage. Computerworld
Microsoft investigated Microsoft 365 Copilot access issues under incident CP1470554. Worth noting for anyone who fields "Copilot is down" reports. CyberSecurityNews

Microsoft / Azure

Microsoft restricted access to its AI tools for users under 13 as part of a new AI safety plan. Türkiye Today

Open-Weight Ecosystem

Released: Baseten added DeepSeek-V4.1-Flash to its Model APIs with a 1M-token context window, giving teams a US-hosted path to run the model without touching DeepSeek's own API. Unite.AI
The open-weight price war is now global, with Zhipu's promo ending and DeepSeek undercutting on price, which pushes hosted-model costs down across the ecosystem. finance.biggo.com

Databricks Watch

AI / ML

GA: Databricks-provided MCP connectors for Genie One and Genie Code connect Google Drive, Gmail, Microsoft 365, Atlassian, Slack, and GitHub, with every tool call authorized and logged through Unity Catalog and Unity Gateway. Per-user OAuth means tokens are not shared. Release Notes, Release Notes
GA: DeepSeek V4.1 Flash is available on Unity Gateway with text and image inputs via Foundation Model APIs; customers remain responsible for compliance with applicable terms. Governance teams should decide their position on this model before users find it. Release Notes

Preview

Preview: The ai_search function can now generate citations showing which retrieved chunks support each answer. Useful for audit trails on AI-assisted analysis. Release Notes

Announced

Announced: In late September 2026, Genie Agents will be restricted to data sources explicitly attached to the agent; instruction-referenced tables will no longer widen access. Review agent instructions now if you rely on that behavior. Release Notes
Announced: The partner-powered AI features toggle can no longer be disabled in the settings UI and will be removed on November 1, 2026; workspaces keep their current value, and later changes require the account team. Record your setting before the toggle disappears. Release Notes, Release Notes
Announced: Databricks published its Adaptive Instructed-Retriever, claiming frontier-quality enterprise search at half the latency. Databricks Blog

Data & Platform

GA: OpenSharing now covers foreign Iceberg tables and foreign schemas/tables federated through Lakehouse Federation, so providers can share external data without copying it into Databricks. Sharing foreign schemas materializes data on the provider side and incurs compute and storage costs. Release Notes, Release Notes
GA: The Databricks Excel Add-in lets users browse Unity Catalog tables, run SQL, build pivot tables with live query, and write data back, all under governance. A big deal for analyst workflows that live in Excel. Release Notes
GA: Lakebase is now PCI-DSS and HITRUST compliant in all AWS regions where it is available, for workspaces with the compliance security profile. HITRUST coverage matters directly for healthcare workloads. Release Notes
GA: The Jobs API now accepts performance_target STANDARD for one-time runs, including through Airflow's DatabricksSubmitRunOperator, a cost lever for non-urgent jobs. Release Notes
GA: Continuous Lakeflow pipelines can now set a one-hour weekly maintenance window for restarts from runtime and infrastructure updates. Release Notes
GA: New time_bucket SQL function aligns timestamps to fixed-width intervals, such as 15-minute buckets, where date_trunc does not fit. Release Notes

Preview

Preview: Zerobus Ingest into tables backed by default storage is in Public Preview in Lakeflow Connect. Release Notes
Preview: External secrets connect a Unity Catalog schema to AWS Secrets Manager as read-only securable objects governed by Unity Catalog privileges. Release Notes
Preview: ABAC DENY policies can block the MANAGE ACCESS CONTROL privilege, including for object owners, and always take precedence over grants. A useful control for separation of duties. Release Notes
Preview: New Lakeflow Connect connectors in Beta ingest data from Anysphere (Cursor) organizations, Celigo audit logs, and Anaplan audit trails. Release Notes, Release Notes, Release Notes

Announced

Announced: In early October 2026, horizontal scaling for Databricks Apps turns on automatically for workspaces with the compliance security profile, running apps across multiple instances behind one URL. Release Notes
Announced: Incremental ingestion for Salesforce formula fields in Lakeflow Connect will soon be GA and will replace full snapshots as the default, cutting cost and run time. Release Notes

Regulatory Radar

US Federal

An Asia Times opinion piece argues the US should adopt "AI poisoning" countermeasures to sabotage Chinese model distillation, following Anthropic's disclosure. No policy has been proposed; watch whether the idea gets official traction, since countermeasures could affect model output quality for everyone. asiatimes.com
OpenAI reportedly hired Senator Chuck Schumer's daughter as AI firms build lobbying ties ahead of the midterms. A signal of how actively labs are working Washington before any federal AI legislation moves. New York Post

Healthcare-specific

The FTC rescinded its 2021 Policy Statement on Breaches by Health Apps and Other Connected Devices. The Health Breach Notification Rule itself still stands, but the withdrawal signals a lighter enforcement posture; teams overseeing member-facing apps should not relax breach practices on this news alone. FTC
CMS announced a crackdown on a $3.4 billion medical equipment supplier fraud scheme in which companies billed Medicare for deceased beneficiaries. Expect continued CMS pressure on DME claims integrity, which touches encounter and claims data quality work at every payer. CMS

Worth Your Time

Health Plans: Your BI Tells You MLR Moved. Can Your AI Tell You Why? A Databricks post aimed squarely at health plan finance and analytics teams, on moving from dashboard readouts to AI that explains medical loss ratio movement. The most directly relevant read of the week for this audience. Databricks Blog
DeepSeek's new model sets a template for powerful LLMs that run lean. The Register's take on why V4.1-Flash matters beyond price: capable models that need less memory change the economics of serving AI. The Register
Rapidly scaling online storage to serve over 1 billion ChatGPT users. OpenAI's engineering story of growing Habitat from a Python library to a platform handling 22 million requests per second. Good systems reading for the data engineers. OpenAI
The 40-year-old database rule agents just broke. Databricks on how agent workloads collapse the OLTP/OLAP split and what an LTAP design looks like. Databricks Blog
Mistral bets enterprise AI will be about control, not just intelligence. TechTarget on Mistral's argument that regulated enterprises will pick deployment control over raw capability, a framing relevant to any compliance-heavy shop. TechTarget