The week leaned hard on things you can actually pick up and use. Meta finally shipped a real coding agent in Muse Code, OpenAI quietly upgraded the model behind everyday ChatGPT and handed free users a much better default, Alibaba pushed a 2.4 trillion parameter Qwen to general availability, and Next.js cut a big 16.3 release. Under all the model news, the security story was the loud one: researchers at Black Hat showed that a single public GitHub issue could turn Claude Code, Gemini CLI, and Codex into remote code execution and credential theft, and CISA piled up four actively exploited flaws in tools developers actually run. On the industry side, Google is reportedly paying $1.5 billion to pull in a tiny reinforcement learning startup, AMD bought its way into custom inference silicon, and the layoffs kept coming for product and engineering teams. Here is everything that mattered.

Top Stories This Week

Meta Ships Muse Code and Muse Spark 1.2, Its First Real Coding Agent -

On August 5, Meta released Muse Code in beta, a terminal and CI coding agent powered by its newest model, Muse Spark 1.2. Muse Code takes on complete software engineering tasks across large repositories, planning changes, writing code, and validating results, and the design detail that stands out is the runtime rather than the model. When a job is big enough it fans out to parallel subagents, each working in its own isolated git worktree so your working copy is never touched, and every model call, tool run, approval, and edit lands in a local append-only event log. That log makes the runtime replay-exact and restart-safe, so a crashed or killed session resumes precisely where it stopped instead of starting over.

For developers, this is worth a real test on long-running work. As TechCrunch reported, Muse Code installs with a single curl command on macOS and Linux, runs either interactively or headlessly through muse exec for scripts and CI, and is available through the Meta Model API. On Meta’s own charts, Muse Spark 1.2 with Muse Code scored 82.9% on Terminal-Bench 2.1, behind Claude Code on Opus 5 at 86.7% but slightly ahead of GPT-5.6 Terra on Codex, and it trailed both rivals on DeepSWE 1.1 at 59.3%. Those are self-reported within-family numbers, so treat the launch chart as a starting point and run your own evals. The honest read is that Meta is late to this fight and its edge is the crash-safe runtime and persistent subagents, not benchmark supremacy, but the cost angle and the restart-safe design make it a credible option for big, multi-step jobs.

OpenAI Retunes GPT-5.6 Sol and Makes GPT-5.6 Luna the Free ChatGPT Default -

On August 6, OpenAI updated ChatGPT with a retuned GPT-5.6 Sol for Plus and Pro users and a new default model for everyone else. The Sol update is chat-focused: more reliable facts, more focused answers, less unnecessary formatting, and a new slider that lets you choose how much reasoning effort the model spends on each response. Free and Go users get GPT-5.6 Luna as their default, replacing GPT-5.5 Instant, which OpenAI says cuts responses with at least one factual error by 62% on financial, medical, and legal prompts. Unlimited text chats and a Think button that escalates harder questions to deeper reasoning are landing for those users the following week, with limits still applying to uploads, images, and other tools.

The part that matters if you build on OpenAI is what did not change. As The Next Web laid out, this is a chat-experience update, not an API change: the GPT-5.6 Sol and GPT-5.6 Luna versions running in Codex and ChatGPT Work are untouched, and there is no new pricing on the API. So your integrations behave exactly as before, but the consumer default your non-technical colleagues use just got meaningfully more accurate, and a model that cost real money a month ago is now the free tier. It is also a clear retention play against Claude and Gemini, since OpenAI is trading away rate limits to keep its billion-plus monthly users from wandering off to free alternatives.

Alibaba Makes Qwen 3.8-Max Generally Available at 2.4 Trillion Parameters -

On August 3, Alibaba officially released Qwen 3.8-Max, the most capable model in the Qwen family so far and the production replacement for the July preview. Built on the Qwen 3.5 architecture, it is a 2.4 trillion parameter sparse mixture-of-experts model with about 95 billion active parameters, a context window up to 1 million tokens, and text, image, and video inputs. Alibaba pitches it at coding, real-world work, research, and long-horizon agentic tasks, and it is live now on QwenCloud with an OpenAI-compatible and Anthropic-compatible API at $2 per million input tokens and $6 per million output tokens.

The genuinely new thing here is openness at this scale. As an independent specs writeup notes, every previous Qwen-Max tier model stayed closed, but Alibaba says it will publish open weights for both Qwen 3.8-Max and a smaller Qwen3.8-27B, with the release scheduled for the following week and pages pointing at around August 10 to 12. Until the checkpoint and a license actually land, treat the open-weight claim as a promise rather than something you can self-host today, and lean on the API if you want to try it now. On the leaderboards it is already competitive, ranking fifth on LMArena’s text board at launch, but as always the launch benchmarks are the vendor’s, so run it against your own coding and agent workloads before committing.

Next.js 16.3 Lands With Instant Navigations and a Lighter Dev Server -

On August 3, Vercel shipped Next.js 16.3, its biggest update since 16.0 last November. Most of the value lands for every existing app with zero code changes: lower dev server memory use, faster rendering, and better runtime performance. On top of that, 16.3 adds Instant Navigations, an opt-in suite that brings client-driven single-page-app responsiveness to Next.js without giving up the server-driven model, plus improved tooling for AI agents and Turbopack support for the Vite-compatible import.meta.glob API so Server Components can hot-reload local files.

This is a clean upgrade worth taking. Run npm install next@latest to move to 16.3, and you get the memory and performance wins immediately; the Instant Navigations tools are opt-in, so you can adopt them route by route. The release also ships experimental features you can flip on today, including a Rust-based React Compiler that runs directly inside Turbopack instead of going through Babel in Node.js, which avoids the extra work of generating and reparsing code. Vercel recommends all apps upgrade, and given the free performance gains for existing projects, that is easy advice to follow once your test suite is green.

Black Hat Shows One GitHub Issue Can Hijack Claude Code, Gemini CLI, and Codex -

At Black Hat USA 2026, Novee Security disclosed that a single public GitHub issue could compromise major AI coding workflows. A public issue opened by an outside user with no repository privileges could turn Claude Code, Gemini CLI, and Codex against the repositories they were meant to maintain. Testing the tools against the vendors’ own repositories and default workflow configurations, the researchers found paths to remote code execution, credential theft, repository modification, and instructions that persisted between separate agent runs. One Claude Code issue, tracked as CVE-2026-54316, abused the agent’s pre-approved access to Hugging Face, and OpenAI’s Codex issue-deduplication workflow ran two agents in the same writable workspace so the first could plant an AGENTS.md file the next agent read as project instructions.

This is the security story to act on this week. Install the corrected Claude Code 2.1.163 and Gemini CLI 0.39.1, then audit any workflow triggered by public issues or pull requests, stop separate agents from sharing a writable directory, and scope tokens down to the minimum each task needs. OpenAI separated its agent runs into different jobs and read-only checkouts within three days, but the researchers stressed that fixing a vendor’s own repository does not automatically fix the same pattern in workflows other organizations copied. The durable lesson is to treat all repository content, including issue and pull request text, as untrusted input the moment an autonomous agent can read it.

Anthropic Makes Claude Code Auto Mode the Default for Pro, Max, and Team -

On August 7, Anthropic announced that auto mode is becoming the default in Claude Code for Pro, Max, and Team plans. Starting August 14, new sessions on those plans will run in auto mode, which enables longer autonomous work and, per Anthropic’s testing, catches more dangerous commands than manual review. If you have already pinned a different default, nothing changes; if you have not set one, you will get an in-product notice and new sessions will start in auto mode. Anthropic also stopped charging Pro, Max, and Team users for the small classifier overhead that auto mode adds per tool call, effective the day of the announcement.

The practical takeaway is to decide now whether you want your default sessions running more autonomously. Auto mode remains opt-in for Claude Enterprise, the Claude API, Claude on AWS and Bedrock, Google Cloud’s Agent Platform, and Microsoft Foundry, with Anthropic planning to flip those to default over the coming month after giving admins time to review. If you run Claude Code across a team and are not ready for more autonomy on August 14, set an explicit default through managed settings before then rather than letting the change roll out on its own.

Kimi K3 Reaches General Availability in GitHub Copilot -

On August 6, GitHub made Moonshot’s Kimi K3 generally available in Copilot, the open-weight model that made headlines last week when a researcher used it to surface zero-days in Redis. GitHub hosts the model on Fireworks AI and is rolling it out to Copilot Pro, Pro+, Max, Business, and Enterprise across VS Code, Visual Studio, the Copilot CLI, JetBrains, Xcode, Eclipse, the cloud agent, and more. It is billed under usage-based pricing at $3 per million input tokens, $15 per million output tokens, and $0.30 per million cached input tokens, and GitHub briefly paused and then resumed the rollout on August 6 while it dealt with a GitHub Actions incident.

The reason to care is cost-effective agentic coding from a strong open-weight model without leaving your existing Copilot setup. If you want to try it, open the model picker once the rollout reaches you, but note two operational details: Kimi K3 is off by default for Copilot Business and Enterprise, so a plan admin has to enable the policy in Copilot settings before anyone in the org can select it, and because it is usage-based it will show up on your metered bill rather than being bundled. For teams already weighing model choice on price and agent quality, it is another serious option in the picker.

Developer Tools & Platforms

Google Cloud Adds Model Routing to API Gateway -

On August 4, Google Cloud launched model routing for API Gateway in public preview, a serverless ingress layer that accepts OpenAI-compatible requests and dynamically routes them to Gemini, Claude, or OpenAI OSS-GPT backends. The point is to stop hardcoding endpoints or running your own proxy: you define routers with a default model and rules that map virtual model names to specific provider backends, so you can send everything to a cheap Gemini model by default and peel specific requests off to Claude or an open-weight GPT. It can run standalone for simple rate limiting and token tracking, or pair with the Gemini Enterprise Agent Platform so your agent’s egress goes through an Agent Gateway for governance before API Gateway handles the routing. For teams tired of maintaining a homegrown LLM proxy, this is a managed way to unify multi-provider traffic behind one edge.

Cursor Adds Google Workspace Plugins for Gmail, Drive, and Calendar -

On August 3, Cursor shipped Google Workspace plugins that give coding agents direct access to Gmail, Google Drive, and Google Calendar, with Docs, Sheets, and Chat also mentioned. The plugins let an agent search and read mail and draft messages, search and create Drive files, and read and update calendar events, all without leaving the editor, so you can have the agent pull a spec from a Doc, write the code, and draft a status email in one flow. Under the hood the plugins connect to Google’s remote Model Context Protocol endpoints, which are still in developer preview, and you install them from the Cursor Marketplace or the Customize page and authorize with a one-time Google login. Because you are handing an agent access to your inbox and files, scope the permissions to the minimum your use case needs and keep to draft-first, human-approved sending until you have tested the behavior on throwaway data.

Django 6.1 Released With Model Field Fetch Modes -

On August 5, the Django team released Django 6.1, with model field fetch modes for configuring on-demand fetching behavior, database-level delete options for ForeignKey.on_delete, and dictionary-based email settings. The fetch modes are the standout for anyone who has fought accidental N+1 queries, since they give you a declarative way to control when related fields are loaded. The release also marks the end of mainstream support for Django 6.0, whose final bug fix and security release, 6.0.8, went out August 4; 6.0 keeps getting security and data-loss fixes until April 2027. You can grab 6.1 from PyPI, and if you are on 6.0 it is worth planning the upgrade before mainstream fixes stop.

ClickHouse 26.7 Speeds Up GROUP BY and Vector Search -

On August 6, ClickHouse shipped its 26.7 release with 61 new features, 112 performance optimizations, and 329 bug fixes. The headline wins are a faster GROUP BY ... ORDER BY ... LIMIT path, three JOIN improvements, faster vector search using a QBit type, position-aware phrase search, and EXPLAIN ANALYZE for actual runtime query plans. It also adds dedicated Remote and RemoteSecure table engines so you can define ClickHouse-to-ClickHouse federation persistently instead of leaning on the remote() table function each time. If you run analytics or vector workloads on ClickHouse, the query-plan visibility from EXPLAIN ANALYZE alone is a reason to read the release notes and plan an upgrade.

Security

CISA Flags Actively Exploited Flaws in TeamCity, Langflow, Tomcat, and N-central -

On August 5, CISA added a batch of actively exploited vulnerabilities to its Known Exploited Vulnerabilities catalog, and they hit tools developers actually run. The most severe is IBM Langflow’s CVE-2026-9198, a critical 9.8 code injection that gives unauthenticated remote code execution on default Langflow deployments and is fixed in 1.10.1. Alongside it, Apache Tomcat’s CVE-2026-34486 (7.5) is an EncryptInterceptor bypass stemming from an incomplete earlier fix, and N-able N-central’s CVE-2026-18556 (8.2) is an authentication bypass whose incomplete first patch spawned a follow-up, CVE-2026-18577. Separately, CISA added JetBrains TeamCity’s CVE-2026-63077, a deserialization bug rated 9.8 that allows unauthenticated remote code execution on on-prem servers, fixed in TeamCity 2026.1.3 and 2025.11.7.

The action is the same for all of them: patch to the fixed release now and check exposure. Federal agencies were given deadlines of August 7 for the Langflow, Tomcat, and N-central bugs and August 8 for TeamCity, which is a reasonable urgency bar for everyone else too. Notably, researchers attributed part of the Tomcat exploitation to an AI-assisted autonomous campaign that used a model through an agent framework to hunt and hit internet-exposed devices, so keeping these systems patched and off the open internet matters more than usual.

Trojanized AI Agent Skills Rack Up 1.7 Million Installs -

On August 7, CSO Online reported on a supply chain attack that trojanized AI agent skills to deploy a credential stealer, uncovered by researchers at Zenity and presented at Black Hat USA 2026. The campaign started July 11, when malicious skills that typosquatted popular AI services were uploaded to the open skills.sh ecosystem, and by August 2 they had amassed more than 1.7 million combined downloads. The skills instructed AI agents to fetch and install a credential stealer straight from GitHub after an earlier attempt to use malicious npm and PyPI packages was blocked, and the collection logic targeted developer workstations, CI runners, and agent workspaces for SSH keys, cloud credentials, Git and package-manager tokens, Kubernetes and Docker configs, and .env files. It is a sharp reminder that shareable agent instruction files are now a supply chain of their own, and Zenity launched a free service called AI Total that detonates a skill inside a sandboxed agent to watch what it actually does before you enable it.

Funding & Industry Deals

Google Reportedly in $1.5B Talks for RL Startup Mechanize -

On August 5, reports emerged that Google is in talks to license technology and hire talent from Mechanize, an AI startup focused on reinforcement learning environments, in a deal worth more than $1.5 billion. Business Insider first reported the discussions, describing a non-exclusive licensing agreement plus the hiring of some of Mechanize’s model evaluation and development experts. The eye-popping part is the trajectory: Mechanize raised just $9.1 million in a seed round on April 24 at a $500 million valuation, so this is roughly a hundred days from a small seed to a reported nine-figure reverse acquihire. It fits a clear Google pattern after similar deals with Character AI and Windsurf, whose former CEO now leads Google’s Antigravity coding platform, and it is another sign that talent and know-how in RL environments are being priced like scarce infrastructure.

AMD Buys Taalas to Hardwire AI Models Into Silicon -

On August 6, AMD agreed to acquire Taalas, a Toronto startup that builds inference chips hardwired for a single AI model rather than general-purpose accelerators. The tradeoff is flexibility for speed and cost: Taalas says its accelerators can produce output for specific models thousands of times faster than a traditional GPU. AMD did not disclose the price, and Taalas had raised $219 million since its 2023 founding. The deal continues AMD’s buying spree to fill out its Helios rack-scale systems, which it has started shipping to customers including Meta and Microsoft, and it lands about seven months after Nvidia paid $20 billion for assets from Groq, underscoring how the GPU leaders now want integrated systems rather than just processors.

HappyRobot Raises $150M Series C for Enterprise AI Agents -

On August 4, HappyRobot raised a $150 million Series C led by Prysm Capital and co-led by Eurazeo, at a $1.2 billion post-money valuation. The company builds agentic AI for supply chains and operations and is expanding from logistics into insurance, energy, telecoms, and airlines. Existing backers a16z, Base10, and Y Combinator doubled down, with strategics including Koch Disruptive Technologies and Deutsche Telekom’s T.Capital joining, bringing total funding to around $200 million. It is another data point that enterprises are willing to pay for AI agents aimed at the messy, high-stakes operational work they cannot afford to get wrong.

Layoffs: Etsy, Google, Salesforce, and Zillow

  • Etsy: On August 5, Etsy said it will lay off about 220 employees, roughly 12% of its workforce, with most cuts concentrated in Product and Engineering, leaving around 1,600 staff. CEO Kruti Patel Goyal framed it as a restructuring rather than a cost-cutting move and pushed back on the idea that AI is replacing people, and it landed six days after Etsy sold Depop to eBay for $1.4 billion.
  • Google: On August 8, a WARN filing revealed Google is cutting 52 positions in Washington state, including software engineers, engineering managers, product managers, UX designers, and recruiters across its Seattle, Kirkland, and Redmond offices, its largest publicly disclosed cut there since 2023.
  • Salesforce: Salesforce is eliminating 133 jobs across California and Washington effective October 5, its third round of cuts this year, weighted toward engineering roles even as CEO Marc Benioff credits AI tools for efficiency gains.
  • Zillow: On August 7, Zillow filed to cut 91 Washington jobs starting in October, concentrated in product managers, software engineers, UX researchers, and machine learning staff, the roles that built its core product.

The Numbers That Matter

  • 2.4 Trillion Parameters in Alibaba’s newly GA Qwen 3.8-Max, about 95 billion active, with open weights promised the following week
  • 82.9% Meta Muse Spark 1.2’s Terminal-Bench 2.1 score with Muse Code, behind Claude Code on Opus 5 (86.7%)
  • 62% Fewer factual errors OpenAI claims for GPT-5.6 Luna over the GPT-5.5 Instant it replaces as the free default
  • 9.8 CVSS score for both the JetBrains TeamCity and IBM Langflow remote code execution bugs CISA flagged as actively exploited
  • 1.7 Million Combined installs of trojanized AI agent skills that shipped a credential stealer
  • $1.5 Billion Reported size of Google’s talks to license technology and hire talent from RL startup Mechanize
  • 220 Etsy employees laid off, about 12% of staff, mostly in Product and Engineering

Quick Hits

  • Qwen 3.8-Max GA - August 3. Alibaba’s 2.4 trillion parameter model goes generally available on QwenCloud, with open weights promised the next week.
  • Next.js 16.3 - August 3. Instant Navigations, a lighter dev server, and an experimental Rust-based React Compiler in Turbopack; upgrade with npm install next@latest.
  • Cursor Google Workspace Plugins - August 3. Coding agents get access to Gmail, Drive, and Calendar through Google’s remote MCP endpoints.
  • Google Cloud API Gateway Model Routing - August 4. A public-preview serverless layer routes OpenAI-compatible requests to Gemini, Claude, or OpenAI OSS-GPT.
  • HappyRobot Series C - August 4. A $150 million raise at a $1.2 billion valuation for enterprise operations agents.
  • Meta Muse Code - August 5. Meta’s first terminal and CI coding agent ships in beta on Muse Spark 1.2 with subagents and a crash-safe event log.
  • Django 6.1 - August 5. Model field fetch modes, database-level ForeignKey.on_delete options, and dict-based email settings; 6.0 hits end of mainstream support.
  • Google Mechanize Talks - August 5. Google reportedly in $1.5 billion talks to license tech and hire from the RL environments startup.
  • Klaviyo Acquires Agency - August 5. Klaviyo buys Elias Torres’ AI customer success startup, with Torres joining as chief product officer.
  • Etsy Layoffs - August 5. About 220 jobs cut, roughly 12% of staff, mostly Product and Engineering, days after selling Depop.
  • CISA KEV Additions - August 5. TeamCity, Langflow, Tomcat, and N-central flaws flagged as actively exploited, with federal deadlines of August 7 and 8.
  • GPT-5.6 August Update - August 6. GPT-5.6 Sol retuned for chat with an effort slider, and GPT-5.6 Luna becomes the free ChatGPT default.
  • Kimi K3 in GitHub Copilot - August 6. Moonshot’s open-weight model reaches general availability under usage-based billing, hosted on Fireworks AI.
  • ClickHouse 26.7 - August 6. Faster GROUP BY and vector search, EXPLAIN ANALYZE, and new Remote table engines.
  • AMD Buys Taalas - August 6. AMD acquires the Toronto startup that hardwires single AI models into inference silicon.
  • Black Hat AI Coding RCE - August 6. Novee shows a public GitHub issue can drive Claude Code, Gemini CLI, and Codex into RCE and credential theft; patched in Claude Code 2.1.163 and Gemini CLI 0.39.1.
  • Trojanized AI Skills - August 7. Zenity details 1.7 million installs of AI agent skills that dropped a credential stealer.
  • Claude Code Auto Mode Default - August 7. Anthropic says auto mode becomes the default for Pro, Max, and Team plans starting August 14.
  • Zillow Layoffs - August 7. 91 Washington jobs cut, concentrated in product and engineering roles.
  • Sarvam AI Trillion-Parameter Plan - August 8. The Bengaluru startup opens a San Francisco office and lays out plans for a one-trillion-parameter model.
  • Google Washington Layoffs - August 8. A WARN filing shows 52 positions cut, including software engineers and engineering managers.

The through line this week was that the frontier is getting cheaper and more agentic at the same time it is getting more dangerous to wire up carelessly. You can now get a strong model for free in ChatGPT, a 2.4 trillion parameter Qwen through an API, a crash-safe coding agent from Meta, and Kimi K3 in your Copilot picker, all in the space of a few days. But the same agentic tools that make those launches exciting are exactly what Black Hat researchers turned into remote code execution with nothing more than a public GitHub issue, and CISA spent the week cataloguing real-world exploitation of the databases, CI servers, and monitoring tools those agents run on top of. If you do only a few things after reading this, patch Claude Code and Gemini CLI, roll out the TeamCity and Langflow fixes, and take a hard look at what your coding agents are allowed to read and run. Next week, watch for whether Alibaba actually publishes the Qwen 3.8-Max open weights it promised, how Muse Code holds up in independent evals, and what OpenAI does with its unlimited free tier once the Think button is live. See you then.