Grok Build was uploading your entire codebase to Google Cloud. Silently. Before anyone asked.

The Verge broke the story yesterday: SpaceXAI's coding tool was caught exfiltrating users' full repos to cloud storage before anyone reported it, at which point SpaceXAI turned it off. No opt-in. No disclosure. A security researcher published the findings and the feature disappeared.

This is not an edge case. This is the default behavior of a product that shipped to real operators writing real production code.

The lesson is not "don't use Grok Build." The lesson is that every AI coding tool touching your repo is making decisions about your data, and most of them are not advertising those decisions. The tool that asks loudest for your codebase context is also the tool with the most interesting incentive to keep it.

Know your data perimeter before you paste. That is the only rule.

Meanwhile: Codex crossed 7M users and added a million in roughly a day, per Latent Space. Claude Code's reporting stayed quiet. That gap is worth watching.

Two moves worth stealing from today's tools before the list. One: point firecrawl at three to five competitor pages on a schedule, diff each new scrape against the last, and let a Claude call flag any pricing or positioning change before your customers ask about it. One API key, an afternoon to wire, no monitoring subscription. Two: when you spawn subagents, send each to the cheapest model that clears the task instead of defaulting everything to your top model, and run the independent ones in parallel. Same work, cheaper and faster.

Today I have 9 drops, 2 tools, and 5 signals.

START HERE. New here? This part is for you. Operators, skip to The Drops.

What is "the cloud"? When people say something runs "in the cloud," they mean it runs on someone else's computer, reached over the internet, instead of on your own machine. Today's opening story is about a tool quietly sending code to the cloud, that is, off your computer and onto a company's servers. Useful, but worth knowing when it happens.

Do this today (2 minutes, no coding): Open whatever AI tool you already use and find its settings. Look for anything labeled data, privacy, or training. You are not changing a thing. You are just learning where a tool tells you what it does with what you give it. That habit is the whole lesson today.

How to use today's picks: Below are 9 projects and 2 tools builders are using right now. You do not need to install any of them to get value. Read the one line pitch, click through, and skim the README (a project's plain-language "what this is and how to start"). You are learning the landscape. Installing comes later, only when one solves a problem you actually have.

Plain English.
model: the AI "brain" itself, like Claude or GPT. It is the thing that was trained to read and generate.
API: a way for two pieces of software to talk to each other automatically. "Using the API" means your app talks to the model directly, with no person clicking buttons.

We read your replies. You told us you are here to learn, so this block is for you. Hit reply with any word you want decoded, and we will define it here.

[Repo] Duix-Avatar, Truly open-source AI avatar and digital human toolkit with 13,993 stars. It handles offline video generation and digital human cloning without a cloud dependency. If you are building a video-presence layer for an agent or a product demo that cannot send data to a third-party API, this is the local alternative.

[Repo] firecrawl, Web scraping and interaction API with 151,021 stars. Gives an agent structured access to the open web: search, scrape, crawl, and interact at scale. The move is wiring it as a tool call in an agent that needs live web context without you managing a browser.

[Repo] langchain, The agent engineering platform at 141,781 stars. If you are assembling multi-step tool chains and want a component-level abstraction rather than writing raw SDK loops, this is the standard toolbox. Know what you are getting: an abstraction with real adoption debt, not a magic layer.

[Repo] llama.cpp, The C++ inference engine that made local models practical. Runs on CPU and GPU, no cloud required. If your pipeline has a latency or privacy constraint that rules out hosted models, this is still the baseline you benchmark against.

[Repo] goose, Multi-provider, MCP-native agent framework from Block, now moving to the Linux Foundation under the AAIF. It carries 70+ extensions and supports custom distributions. A credible peer to the Claude Code stack for operators who want provider flexibility without framework lock.

[Repo] awesome-ai-agents-2026, The most comprehensive curated list of AI agents, frameworks, and tools in 2026: 300+ resources across 20+ categories, updated monthly. Useful when you need to audit the field before committing to a framework or find a category of tool you know exists but cannot name.

[Repo] awesome-agents, Kyrolabs' curated list of AI agents at 2,622 stars. Narrower and more opinionated than the above. Good for a second pass when you want a shorter, more filtered starting point.

[Skill] self-learning-skills, A self-improving skill for AI coding agents including Claude Code and Cursor. It watches your session, recognizes a hard-won path, and harvests it into a reusable skill for next time. The 864-star count understates what this solves: the problem is not building skills, it is remembering to build them when you are already deep in flow.

[Repo] ccstatusline, A customizable status line for the Claude Code CLI at 11,730 stars. Powerline support, themes, and real-time session state. Not cosmetic: a readable status line is the difference between knowing what your agent just did and guessing.

[Affiliate] Lindy, a no-code platform for building AI agents that actually do the work: triage the inbox, qualify leads, take meeting notes, run multi-step flows across your tools. Unlike if-this-then-that automation, its agents use an LLM to make judgment calls in context.

[MCP] ditto, Mines your Claude Code and Codex session logs and builds a local you.md agent profile from them. The non-obvious use: it accumulates the operator-specific patterns your agent keeps relearning across sessions and surfaces them as a persistent context file. Wire it once, and your agent starts the next session already knowing your preferences. The gotcha: it reads your logs, so be deliberate about which sessions you expose it to.

[Repo] pipecat, Open-source framework for voice and multimodal conversational AI at 13,423 stars. If you are building a voice-first agent layer, this handles the real-time audio pipeline so you are not managing WebRTC and VAD from scratch. One config trick: set the transport to DailyTransport for the fastest time-to-working-demo on new hardware.

- Grok Build exfiltrated user codebases to Google Cloud before anyone noticed. SpaceXAI's coding tool was silently uploading full repos; a researcher published findings and the feature was turned off. The operator consequence: any AI coding tool with repo access is making data decisions you need to audit before you trust it with production code. (The Verge)

- Codex crossed 7M users, adding roughly 1M in about a day. Usage is up more than 10x in six months. Claude Code's public reporting stayed quiet during the same window. If Codex is your competition's primary coding tool and you are building on Claude Code, that gap in visibility is worth tracking. (Latent Space)

- Node.js 20 is being deprecated on Vercel on October 1, 2026. Node 20 hit end of life on April 30. If your agents or pipelines deploy any Functions or Builds on Vercel still pinned to Node 20, you have until October 1 to migrate or they break. Check which projects are affected in your Vercel dashboard now, not in September. (Vercel)

- iOS 27 public beta is live, with Apple's revamped Siri open to everyone. If you are building anything on Apple's on-device AI layer, the public beta is the first real test surface outside the developer program. The relevant question is not Siri itself but which on-device model capabilities are now accessible to third-party apps. (TechCrunch)

- OpenAI is positioning Codex as a super-app, potentially replacing ChatGPT chat as the primary interface. Stratechery's read is that OpenAI is refashioning Codex as the new core product and may be walking back from the chat category it invented. If that shift lands, the agent-first interface becomes the default, and every workflow built on a chat-first assumption gets revisited. (Stratechery)

We build The AIgent's engine in the open. An honest look at what we are making, what broke, and where it is headed. This week: the publish gate that caught a fabricated quote in our own draft.

AI-drafted content fails confident. It will hand you a specific, plausible, attributed quote that was simply never said.

Our daily issue is drafted by our own system, and every draft passes a verification gate before it can ship. This week the gate earned its keep twice in one issue. The draft quoted a well-known investor calling a new model a "step change." Checking the primary source, his actual post said something related but materially different, and the phrase "step change" appears nowhere. Same pass, second catch: a pricing comparison attributed to a vendor's changelog that the changelog does not make.

Neither read as suspicious. That is the point. Fabrications do not arrive looking wrong; they arrive looking like the most quotable sentence in the piece.

The gate works by claim class, because different claims fail differently. Direct quotes: find the primary source and match it verbatim, not approximately. Numbers: re-pull them live at publish time, never trust the draft's copy. For this issue that meant re-checking eleven repositories and seven star counts against the live API; six had drifted. Comparative claims, cheaper, faster, better-than: the comparison must appear in the cited source itself, or it goes. And anything that survives none of those gets paraphrased down to what you can actually stand behind.

One issue, two fabrications, zero shipped. The uncomfortable takeaway for anyone publishing with AI in the loop: if you do not have a claim-class gate, you have already shipped a quote that does not exist. You just have not found it yet.

Publishing anything AI-drafted? Hit reply and tell us your verification step, or admit you don't have one yet. We read every one.

The Grok Build story is the one I keep returning to today. Not because SpaceXAI is uniquely untrustworthy, but because "silently uploading your codebase" is a product decision that shipped, passed review, and reached real users before a researcher caught it. The data perimeter on your agent stack is not a setting. It is a thing you have to actively own. Check it before you need to.

See you Thursday.