Presented by
Good morning, AI enthusiasts. A security researcher downloaded Anthropic's entire Claude Code codebase, not from a hack, but from an npm package that shipped with a full source map attached.
What's inside reveals where AI coding tools are heading, with persistent background agents, voice commands, browser control, and features Anthropic never announced.
In today's recap:
Anthropic's accidental npm leak, 44 hidden Claude Code features exposed
Salesforce turns Slack into a universal AI agent router
Automate team standups and task routing with Slack Agentforce
OpenAI closes $122B round at an $852B valuation
4 new AI tools, prompts, and more
ANTHROPIC
Anthropic's Claude Code source code accidentally leaked
Chaofan Shou/X
Recaply: Anthropic just exposed Claude Code's entire source code through an npm source map in v2.1.88, revealing 512,000 lines of TypeScript, 44 feature flags, and an unreleased roadmap it never meant to publish.
Key details:
Bun's bundler included a 59.8MB source map in the npm package by default, giving security researcher Chaofan Shou access to the full codebase, which he archived to GitHub within hours.
The leak covers 1,900 TypeScript files and 44 hidden features not yet shipped, including KAIROS, a 24/7 background agent, voice commands, and full browser control via Playwright.
The leaked code shows Undercover Mode removes AI traces from Anthropic employee commits to public repos, with no remote off switch.
Anthropic pulled v2.1.88 from npm within hours, but the archive was already live on GitHub with 1,100 stars. This is Anthropic's second source exposure in five days.
Why it matters: This leak isn't just embarrassing for Anthropic. It hands competitors a clear picture of where Claude Code is heading. The biggest reveal is KAIROS, an agent designed to run in the background, pause on its own, and resume without any user input. That points to Anthropic building toward persistent AI that works while you're away from your desk. The mistake behind it, a source map left inside a production npm package, is the kind of build configuration error any fast-moving team could make.
PRESENTED BY MINTLIFY
The fastest-growing repo on GitHub is a one person team!
OpenClaw went from 9K to 185K GitHub stars in 60 days — the fastest-growing repo in history.
Their docs? One person, plus Claude. They scaled to the top 1% of all Mintlify sites, shipping 24 documentation updates a day.
SALESFORCE
Salesforce turns Slack into enterprise AI control layer
Salesforce
Recaply: Salesforce just shipped a full agent orchestration layer inside Slack, turning Slackbot into a universal router that can build, chain, and trigger AI agents directly from within team workflows.
Key details:
Slackbot can now route tasks to any Agentforce agent or MCP-connected third-party tool. Teams ask it a question, and it finds the right system and runs the job automatically.
The update also ships desktop computer vision, deep research mode, and AI skills sharing, all included for paid Slack customers.
Slack says the desktop feature doesn't take control of your mouse or open apps outside the workspace. It handles tasks in the background while users stay in control.
The feature is live now for existing paid plans, alongside voice input, recurring automations, and a built-in CRM for small businesses.
Why it matters: Slack was already where most enterprise conversations happen. Now it's where the actual work gets done. Salesforce is betting that companies won't adopt separate AI tools if Slackbot can route every request to the right agent in the background. For competitors like Microsoft Teams, this is a direct challenge. Slack's AI play isn't a new application layer. It's a takeover of the coordination layer most enterprise teams never quit.
TUTORIAL
Automate team stand-ups and task routing with Slack Agentforce
Recaply: In this tutorial, you will learn how to set up Slack agent workflows that handle your team's recurring coordination tasks, from daily standups to ticket routing and meeting follow-ups, without manual work.
Step-by-step:
Open Slack and go to your workspace → Automations, then select "Create workflow." Choose "Slackbot" as the trigger source and pick a recurring schedule, such as weekdays at 9 a.m., to define when the workflow fires.
Add a "Slackbot: Collect responses" action and type your standup prompt, for example: "What did you complete yesterday? What are you working on today? Any blockers?" Slackbot will send this to each team member on schedule.
Add a "Slackbot: Summarize and post" step to collect all replies and post a formatted summary to your team channel. Set the destination to your project channel and enable "Auto-tag open blockers" if the option appears.
Add a routing step by selecting "Slackbot: Route task" and linking it to your project management app, such as Linear, Jira, or Asana. Slackbot will read the blockers from step 3 and create or update tickets in the linked tool.
Test the workflow by clicking "Run now," review the standup summary in your channel, and confirm that any blockers were logged in your connected project tool. Activate the schedule when ready.
Pro tip: Use the AI Skills feature in Slack to save your standup workflow as a reusable skill and share it across the company so other teams can turn it on with one click.
TOOLS
Trending AI Tools
🤖 1-Bit Bonsai - PrismML's first commercially viable 1-bit LLM
⚙️ JSSE - A JavaScript engine built entirely by Claude Code in YOLO mode
🎥 Veo 3.1 Lite - Google's most cost-effective video generation model
💻 KAT-Coder-Pro V2 - KwaiKAT's latest agentic coding model
NEWS
What Matters in AI Right Now?
OpenAI closed a record-breaking $122B funding round at an $852B valuation, with SoftBank, Andreessen Horowitz, and D.E. Shaw among the investors. The company is now generating $2B in monthly revenue but remains unprofitable as IPO pressure builds.
Google introduced Veo 3.1 Lite, its most cost-effective AI video model, now in paid preview via the Gemini API and Google AI Studio. It supports text-to-video and image-to-video at 720p or 1080p, priced at roughly half the cost of Veo 3.1 Fast.
Oracle is cutting thousands of employees to fund its AI data center expansion, with TD Cowen estimating the reductions could reach 20,000 to 30,000 workers, roughly 18% of its 162,000-person workforce.
H Company launched Holo3, a computer use model that scored 78.85% on the OSWorld-Verified benchmark for desktop agents. With 122B total parameters but only 10B active, it's designed to outperform larger proprietary models at lower cost.
Starcloud raised $170M in a Series A led by Benchmark and EQT Ventures, reaching a $1.1B valuation just 17 months after Y Combinator demo day. The company is building orbital data centers with Nvidia Blackwell chips, targeting cost parity with ground-based systems once Starship launches commercially.
Northwestern University researchers found that 60% of U.S. federal judges now use at least one AI tool in their judicial work, based on a random-sample study of 502 judges. Legal research leads at 30% usage, though fewer than half of respondents said their courts provide any AI training.
OkCupid and Match Group settled with the FTC over sharing nearly 3 million user photos with facial recognition company Clarifai in 2014 without user consent. The companies paid no fine, agreeing instead to a permanent ban on misrepresenting how they collect and share personal data.
🧡 Enjoyed this issue?
🤝 Recommend our newsletter or leave a feedback.
How'd you like today's newsletter?
Cheers, Jason
