ai-powered-markdown-translatorTranslated article from fr to en with gpt-5.4-mini.
A dense week for the AI ecosystem: Alibaba launches its flagship model Qwen3.7-Max designed for the agentic era, GitHub reveals a security incident involving a malicious VS Code extension, NVIDIA publishes an open governance framework for AI agent skills, and Runway renews its video editing model with Aleph 2.0 and Edit Studio. On the developer tools side, Claude Code 2.1.146 renames /simplify to /code-review and Codex 26.519 introduces Appshots on macOS.
Qwen3.7-Max โ Alibabaโs flagship for the agentic era
May 21 โ Alibaba launches Qwen3.7-Max, its new flagship proprietary model, designed specifically for autonomous agents. The model follows the Qwen3.7-Max-Preview and Qwen3.7-Plus-Preview previews evaluated on Arena on May 18.
The positioning is explicitly โAgent Eraโ: Qwen3.7-Max targets tasks that go beyond point-in-time understanding and move into the realm of long-term autonomy and complex orchestration.
| Metric | Value |
|---|---|
| AAII score | 56.6 |
| Progression vs previous | +4.8 pts |
| Continuous autonomy hours (core test) | 35 h |
| Tool calls (core test) | 1,000+ |
| Main tweet views | ~325,000 |
Five announced capabilities:
- End-to-end coding agent โ frontend prototypes, multi-file refactors, real debugging
- Desktop productivity โ tasks via MCP integrations and multi-agent management
- Long-term autonomy โ 35 consecutive hours on a kernel optimization task, 1,000+ tool calls without critical hallucination
- Scaffold-agnostic โ compatible with Claude Code, OpenClaw, Qwen Code, or any custom stack
- Enhanced scientific reasoning โ AAII score: 56.6 (+4.8 pts)
According to Tongyi Lab (Alibaba), Qwen3.7-Max is โour latest proprietary model designed specifically for the age of agentsโ, with a frontier coding agent, a desktop productivity assistant, and long-term autonomy โ reliable every time, scaffold-agnostic.
The model is available via Qwen Studio (chat.qwen.ai) and the Alibaba Cloud API.
๐ Official Qwen3.7-Max blog
GitHub โ Security incident: malicious VS Code extension
May 20 โ GitHub reveals a major security incident: an employee installed a malicious VS Code extension from the Microsoft marketplace, allowing unauthorized access to GitHub internal repositories.
The initial announcement on X generated 13.4 million views, 8,000 reposts and 25,000 likes โ a sign of considerable community impact. GitHub published a 5-part thread detailing the timeline and the measures taken.
โ1/ Yesterday we detected and contained a compromise of an employee device involving a poisoned VS Code extension. We removed the malicious extension version, isolated the endpoint, and began incident response immediately.โ โ Yesterday, we detected and contained the compromise of an employee device involving a poisoned VS Code extension. We removed the malicious version of the extension, isolated the endpoint, and immediately launched incident response.
What happened:
- Vector: malicious VS Code extension from the Microsoft marketplace
- GitHub employee device compromised
- Unauthorized access to GitHub internal repositories
What was not affected (according to GitHub):
- No confirmed impact on customer data (companies, organizations, public or private repositories)
Immediate actions:
- Removal of the malicious extension
- Isolation of the compromised endpoint
- Log analysis and secrets rotation
- Increased infrastructure monitoring
GitHub committed to publishing a full report at the end of the investigation. This incident reignites the debate on the security of the VS Code extension marketplace, which hosts tens of thousands of extensions whose verification remains limited.
NVIDIA Verified Agent Skills โ An open standard for AI agent governance
May 21 โ NVIDIA launches NVIDIA-Verified Agent Skills, an open governance framework to secure AI agent capabilities (skills). The initiative addresses a growing problem: skills expand agent capabilities, but also introduce potential vulnerabilities.
A skill โverifiedโ by NVIDIA rests on four pillars:
| Component | Description |
|---|---|
| Cataloging | Daily synchronization from NVIDIA product teams |
| SkillSpector | Risk scanner based on OWASP LLM and MITRE ATLAS |
| Cryptographic signature | Detached signature skill.oms.sig verifiable via OpenSSF Model Signing |
| Skill Card (SKILLCARD.yaml) | Machine-readable metadata: provenance, dependencies, limitations, status |
The framework detects in particular: vulnerable dependencies, prompt injection, trigger abuse, excessive agency, and tool poisoning.
The technical specification is open via agentskills.io (SKILL.md format), natively compatible with Claude Code, OpenAI Codex and Cursor. NVIDIAโs public skills repository is available on github.com/nvidia/skills.
โVerified means cataloged, scanned, signed, and documented with a skill card.โ โ Verified means cataloged, scanned, signed, and documented with a skills sheet.
The fact that the specification is open and compatible with the three main AI coding tools (Claude Code, Codex, Cursor) suggests an ambition for an industry standard, not just an internal NVIDIA tool.
Runway Aleph 2.0 and Edit Studio โ Video editing up to 30 seconds in 1080p
May 21 โ Runway launches Aleph 2.0, a major update to its video editing model, accompanied by Edit Studio, a new interface dedicated to marketing teams and filmmakers.
The central innovation of Aleph 2.0 is image-level control: the user provides an edited frame as a visual reference, and the model applies the modification consistently across the entire video. Unlike previous approaches where the final result was unknown before generation, this method offers predictable control over the output.
| Feature | Detail |
|---|---|
| Max duration | 30 seconds |
| Resolution | 1080p |
| Control | Reference image frame (image-level) |
| Multi-shot editing | Yes โ automatic application across multiple shots |
| Access | All paid Runway plans |
Aleph 2.0 editing capabilities:
- Precise localized edits: only the modified area changes, the rest is preserved
- Multi-shot editing: changes applied to videos with multiple cuts
- Clips up to 30s in 1080p โ sufficient length for ads and short-form content
Edit Studio use cases: product swap, background change, seasonal campaign variation, lighting correction, removal of unwanted elements, complete restyling.
โWith Aleph 2.0, your edited frame defines what the change should look like, and the model carries it through to the right parts of your video.โ โ With Aleph 2.0, your edited image defines the desired look, and the model applies it to the corresponding parts of your video.
๐ Official Runway announcement
Claude Code 2.1.146 โ /code-review and background agents fixes
May 21 โ Anthropic releases version 2.1.146 of Claude Code with one main new feature and several important stability fixes.
The /simplify command is renamed /code-review, accompanied by an optional effort level (/code-review high). This change explicitly repositions the feature as a code review tool in agentic workflows.
| Area | Fix |
|---|---|
| Windows PowerShell | Regression v2.1.124 fixed for pwsh via winget/Microsoft Store |
| MCP pagination | resources/list and prompts/list no longer lose items beyond page 1 |
| Background sessions | Windows Terminal strobing, NTFS worktrees, /background and claude agents blockages |
| Auto mode | AskUserQuestion preserved when the skill or user depends on it |
| Bedrock | Cross-account auth fixed with awsCredentialExport |
| VSCode Voice | Silent microphone feedback, sox suggestions for WSLg |
Claude Code /usage โ Breakdown of token consumption by component
May 21 โ Boris Cherny (lead Claude Code at Anthropic) announces on X a significant extension to the /usage command: the next version will make it possible to see exactly which Skills, Agents, MCPs and Plugins are consuming tokens in a session.
This feature goes beyond yesterdayโs UI redesign. It adds per-component granularity โ making it possible to identify the most context-hungry extensions. It is already available in CLI and will arrive in the Desktop app soon. The announcement generated 137,527 views on X, with many requests to selectively disable unused connectors.
โIn the next version of Claude Code: run /usage to see a breakdown of which Skills, Agents, MCPs, and Plugins are using your tokens CLI today, coming to Desktop nextโ โ In the next version of Claude Code: run
/usageto see the breakdown of token consumption by Skills, Agents, MCPs, and Plugins โ available in CLI today, on Desktop soon.
Google I/O 2026 โ The official โ100 announcementsโ recap
May 20 โ Google publishes its official recap article โ100 things announced at I/O 2026โ, a reference document for all event launches. Among the confirmed announcements: Antigravity 2.0 CLI (successor to Gemini CLI), Universal Cart (multi-retailer shopping), Google Pics (smart photo management), Ask YouTube (querying YouTube videos), Android XR glasses and the expansion of SynthID. This document consolidates the scope of Google I/O 2026 and serves as a primary source for developers wanting all the official details.
๐ Google blog โ 100 I/O 2026 announcements
Gemini Connected Apps โ OpenTable, Canva and Instacart
May 21 โ Google Gemini expands its Connected Apps to three new partners: OpenTable (restaurant reservations), Canva (visual creation) and Instacart (grocery shopping). Gemini no longer just looks up information โ it takes concrete actions in the userโs digital ecosystem on their behalf. This post-I/O announcement concretely reflects Gemini Sparkโs agentic vision: the agent acts on popular third-party services without leaving the app.
๐ Gemini Connected Apps announcement
GitHub Copilot โ Four developer updates
Metrics reports migrated to copilot-reports.github.com
May 20 โ The download URLs for Copilot Usage Metrics reports migrate from the old Azure Front Door domain (azurefd.net) to the stable domain copilot-reports.github.com. Action required for organizations with firewalls or proxies: add the new domain to the enterprise allowlist as soon as possible. The old domain remains active during the transition period but will be deprecated later. For GitHub Enterprise Cloud instances, the new domain follows the format copilot-reports.SUBDOMAIN.ghe.com. A fallback to *.blob.core.windows.net is planned in case Azure is unavailable.
Cloud agent โ Claude Haiku 4.5 and GPT-5.4-mini at 0.33x
May 18 โ The Copilot cloud agent integrates two new economical models: Claude Haiku 4.5 and GPT-5.4-mini, both billed at 0.33x the standard multiplier (versus 1x for standard models). This expansion makes it possible to choose the right tool depending on the task โ a lightweight, fast model for simple lint fixes or unit tests, a more powerful model for complex refactors. The model is chosen when delegating a task to the cloud agent, making it possible to optimize the cost/quality ratio according to the actual complexity of the work.
Web chat โ Contextual side panel generally available
May 18 โ Copilot Chat on github.com now opens in a side panel directly next to the resource being viewed (PR, issue, repository), without having to navigate to a dedicated page. Context accumulates automatically as you move between multiple PRs and issues in the same session โ cross-resource references remain accessible without losing the thread of the conversation. The full-page immersive experience remains available via โMore โ In immersive chat.โ It is also possible to turn the conversation into an agent session to create PRs or start a deep research task. Available in GA for all Copilot plans.
๐ Copilot Chat web changelog
Genspark integrates Gemini 3.5 Flash for free
May 20 โ Genspark integrates Gemini 3.5 Flash the day after its announcement at Google I/O, available for free in AI Chat Agent and Genspark Claw. The platform (70+ simultaneous models available at once) illustrates its reactive integration strategy: adopt the newest models without forcing the user to choose manually. Gemini 3.5 Flash is Googleโs most capable Flash model to date, optimized for response speed on complex tasks.
Kling AI at Cannes โ RAPHAEL RAPHAEL, 100% AI film
May 21 โ Kling AI presents RAPHAEL RAPHAEL at the 2026 Cannes Film Festival, a feature-film project 100% generated by AI at scale. This announcement is separate from the May 20 post on House of David (the first Hollywood series to integrate Kling at industrial scale, 44 million viewers, #1 on Prime Video US). Kling AI thus consolidates a multi-project presence at Cannes 2026, asserting its position as a major AI video player for cinema.
BFL FLUX Erase โ Object removal and text erasure
May 21 โ Black Forest Labs launches FLUX Erase, an intelligent erasing feature for FLUX images. Three main capabilities: object removal (clean removal with coherent background reconstruction), text erasure (removal while preserving the underlying design) and automatic shadow removal linked to erased objects. The feature is designed for precision retouching (precision inpainting): only the targeted area is modified, without degrading the rest of the image. Available via the web tool flux-tools.bfl.ai/erase and the BFL API (docs.bfl.ai/flux_erase), with immediate access and no waitlist.
๐ BFL announcement
ElevenReader โ 200,000+ premium titles with voice of choice
May 21 โ ElevenLabs integrates 200,000+ premium titles into ElevenReader: best-selling audiobooks and licensed eBooks, including thousands available with the narratorโs voice of choice among ElevenLabs voices. This expansion positions ElevenReader as a customizable audio reading platform, where the user is no longer limited to a bookโs original narrator: they can choose the voice they prefer for any compatible title. The app is accessible at elevenreader.io with free access to get started.
xAI Grok ร OpenCode โ Grok Build in the code CLI
May 21 โ Grok or X Premium subscribers can now use the Grok Build model in OpenCode, a command-line code editor (CLI coding editor) specialized in codebase intelligence. This integration follows the one announced in OpenClaw two days earlier โ xAI is gradually extending its subscription to more third-party tools. Access is included in the existing Grok/X Premium subscription, at no extra cost. OpenCode stands out from OpenClaw for its specific focus on code: codebase analysis, navigation, and assisted refactoring.
๐ xAI announcement
Codex 26.519 โ Appshots, Goal mode GA, locked computer use, plugin sharing
May 21 โ OpenAI releases Codex 26.519 with several new macOS and enterprise features. Appshots let you send the foreground application window into a Codex thread via a double press on both Command keys โ the visible window and the text accessible through the macOS accessibility API are captured. Goal mode exits experimental status and moves to general availability (app, IDE extension, CLI). Remote computer use: Codex can now use desktop applications after the Mac is locked. Plugin sharing is available for ChatGPT Business.
| Feature | Status | Platforms |
|---|---|---|
| Appshots | New (GA) | macOS only |
| Goal mode | GA (was experimental) | App, IDE extension, CLI |
| Remote computer use (locked Mac) | New | macOS + Codex Mobile |
| Plugin sharing | New (Business) | ChatGPT Business |
๐ Codex changelog
Briefs
- Gemini Daily Brief โ general availability โ The daily summary is now available to all Google AI Plus, Pro, and Ultra subscribers (18+), initially in the United States. ๐ source
- Grok ร Imagine โ video game assets (8.1M views) โ Viral demonstration of prototyping game assets directly with Grok and @imagine, with strong gaming community engagement. ๐ source
- Cohere Command A+ โ W4A4 weights available on HuggingFace โ Quantized W4A4 weights for Command A+ are downloadable, enabling deployment on 2ร H100 with +47% speed vs BF16. ๐ source
What this means
The agentic era is taking shape on several fronts at once. The launch of Qwen3.7-Max with 35 hours of continuous autonomy and 1,000+ tool calls, combined with NVIDIA Verified Agent Skills that standardize governance for agent capabilities, signals that the industry is moving from an experimental phase to a structuring phase. The fact that NVIDIA bases its standard on the open spec agentskills.io โ compatible with Claude Code, Codex, and Cursor โ suggests a convergence toward common interfaces for AI agents, regardless of the underlying model.
The security of the AI tool chain is becoming a critical issue. The GitHub incident, in which a malicious VS Code extension from Microsoftโs official marketplace compromised an internal device, illustrates a new attack surface that the industry has not yet fully addressed. The timing is telling: NVIDIA launches its Verified Agent Skills on the same day GitHub announces its incident. Unverified skills and extensions represent a growing attack vector as AI agents take control of critical systems.
Developer tooling is consolidating around consumption transparency. The announcement of Claude Code /usage with component-level breakdown, combined with Codex Appshots and GA Goal mode, reveals a trend: next-generation AI developer tools must now explain what they consume and how. Granularity by Skills/Agents/MCPs/Plugins responds to a concrete need from teams deploying agents in production and wanting to understand the real costs.
Generative video AI is entering the film industry. Runway Aleph 2.0 with its predictable image-level control, Kling AI at Cannes with a 100% AI feature film (RAPHAEL RAPHAEL), and confirmation that House of David (44M viewers, #1 Prime Video US) uses Kling at industrial scale mark a turning point. Studios are no longer testing the technology โ they are integrating it into their standard production pipelines.
Sources
- Qwen3.7-Max โ Alibaba Qwen Tweet
- Qwen3.7-Max โ Official blog
- GitHub โ Security incident (initial announcement)
- GitHub โ Security incident (detailed thread)
- NVIDIA Verified Agent Skills โ Developer blog
- NVIDIA Verified Agent Skills โ Tweet
- Runway Aleph 2.0 โ Official announcement
- Runway Aleph 2.0 โ Tweet
- Claude Code CHANGELOG
- Claude Code /usage breakdown โ Boris Cherny
- Google I/O 2026 โ 100 announcements
- Gemini Connected Apps โ OpenTable, Canva, Instacart
- GitHub Copilot โ Report URL migration
- GitHub Copilot โ Claude Haiku 4.5 and GPT-5.4-mini
- GitHub Copilot Chat web โ GA side panel
- Genspark โ Gemini 3.5 Flash integrated
- Kling AI โ RAPHAEL RAPHAEL in Cannes
- BFL FLUX Erase
- ElevenReader โ 200,000+ titles
- xAI Grok ร OpenCode
- Codex 26.519 โ Appshots
- Codex 26.519 โ Changelog
- Gemini Daily Brief GA
- Grok ร Imagine โ video game assets
- Cohere Command A+ โ W4A4 HuggingFace