Get tomorrow's brief in your inbox
Today: Anthropic shipped auto mode for Claude Code that lets AI decide which actions are safe to run on its own. OpenAI shuttered its TikTok-like Sora app after six months of creepy deepfakes and declining downloads. Arm released its first in-house chip, competing directly with its own licensing partners.
Anthropic launched "auto mode" for Claude Code, fundamentally changing how AI coding assistants work. Instead of requiring permission for every action or running unchecked with --dangerously-skip-permissions, Claude now decides on its own which actions are safe to execute immediately and which need human approval.
The system uses Claude Sonnet 4.6 as a classifier to review each action before it runs, checking for risky behavior like scope escalation, unrecognized infrastructure targets, or signs of prompt injection attacks. It ships with extensive default filters: it allows things like local file operations within project scope, read-only API calls, and installing dependencies already declared in manifest files. It blocks destructive operations like force-pushing to remote branches, downloading and executing external code, and cloud storage mass deletes.
For small businesses using AI coding tools, this matters because it removes the constant interruption of permission prompts without the anxiety of letting AI run wild. If you're using Claude Code to automate routine development tasks, auto mode means fewer clicks and faster workflows. The catch is that AI-based safety filters are probabilistic, not deterministic, so they might still allow risky actions through if user intent is unclear. Anthropic recommends using auto mode in isolated environments, not production systems. If you're a Claude Enterprise or API user, this rolls out over the next few days. Read more
Talat: Local-Only AI Meeting Notes If you're tired of sending your voice recordings to cloud services, Talat offers a privacy-focused alternative to tools like Granola. This Mac app records and transcribes meeting audio entirely on your machine using Apple's Neural Engine. No accounts, no subscriptions, no data leaving your computer. It's a 20MB one-time purchase that captures audio from Zoom, Teams, and Meet, assigns speakers in real time, and generates summaries with action items using a local LLM. You can configure it to use your own cloud LLM if you want, or point it at Ollama for fully local processing. If client confidentiality matters to your business, this is worth checking out. The app costs a one-time fee with no ongoing charges. Read more
Zapier MCP for Microsoft Copilot Zapier shipped Model Context Protocol (MCP) support for Microsoft Copilot, giving it governed access to 8,000+ apps and 30,000+ actions. This means Copilot can now pull action items from OneNote and create tasks in Notion, grab project updates from Coda and draft status emails in Outlook, or research prospects and log findings to your CRM. The setup takes minutes: connect your Zapier account, add the tools you want, and Copilot can work across your entire tech stack. If you're already paying for Microsoft 365, this extends Copilot beyond Microsoft's ecosystem without switching platforms. Free to set up through your Zapier account. Read more
ChatGPT Shopping Interface Update OpenAI updated ChatGPT's shopping interface with side-by-side product comparisons, improved speed, and more up-to-date product coverage. This comes as OpenAI walks away from its built-in checkout feature after disappointing sales numbers. Instead of processing transactions directly, ChatGPT now focuses on visual product discovery: pricing, reviews, and features displayed in a cleaner layout. If you use ChatGPT to research products before buying, the new interface makes comparisons easier. The update is rolling out to all ChatGPT tiers this week. Read more
Spotify Artist Profile Protection Spotify launched a beta feature that lets artists review and approve releases before they appear on their profiles. This addresses the flood of AI-generated tracks being misattributed to real artists due to metadata errors or malicious uploads. Artists in the beta receive email notifications when music is delivered to Spotify with their name attached, then approve or decline it. Only approved releases contribute to stats and show up in recommendations. If you're an independent artist dealing with AI slop landing on your profile, this gives you control. The feature is rolling out to select artists now. Read more
Zapier Marketing Calendar Templates with AI Zapier released marketing calendar templates that combine automation with AI. The social media calendar template uses Zapier Forms for content input, stores posts in Zapier Tables, and automatically publishes to Instagram and Facebook at scheduled times. Add an AI step to the workflow and it can summarize or categorize content before it hits your channels. If you're running social media for a small business, this replaces manual scheduling spreadsheets with an automated system. Free with a Zapier account. Read more
Multi-Agent Systems for Complex Workflows Zapier published a guide on multi-agent systems: multiple specialized AI agents that work together on complex tasks. Instead of asking one agent to handle a multi-step workflow (which increases the chance of errors), you create specialized agents that pass outputs to each other. Example: one agent extracts feature requests from customer calls, sends them to a second agent that generates a product requirements document, then returns the completed doc to a human. This matters for small businesses because it makes AI more reliable for workflows that require multiple steps. If you're already using Zapier AI agents, this explains how to chain them together. Read more
Gap Partners with Google Gemini for AI Checkout Gap became the first major fashion company to offer checkout directly within Google's Gemini AI platform. Shoppers asking Gemini for style advice can buy Gap, Old Navy, Banana Republic, and Athleta products without leaving the chatbot, paying via Google Pay while Gap handles shipping. This signals a shift from traditional search to conversational AI for product discovery. If you sell consumer products, this is the early shape of AI-native commerce: people ask questions, AI recommends products, transactions happen inline. The Gap integration goes live soon. Read more
OpenAI Shuts Down Instant Checkout OpenAI confirmed it's walking away from Instant Checkout, the feature that let users buy products directly through ChatGPT. Walmart reported "disappointing" sales numbers, and OpenAI is now shifting focus to improving product presentation instead of processing transactions. If you were building integrations around ChatGPT checkout, the play now is standalone retailer apps within ChatGPT. OpenAI is working with partners on that approach. Read more
Package Manager Dependency Cooldowns Following the LiteLLM supply chain attack, multiple package managers added "dependency cooldown" features that delay installation of newly updated packages by a few days. This gives the security community time to spot compromised releases before they hit your system. Tools with cooldown support: pnpm, Yarn, Bun, Deno, uv, pip, and npm. If you manage software dependencies, configuring cooldowns is a low-cost security layer. Check your package manager's docs for minimumReleaseAge settings. Read more
Arm Releases First In-House Chip, Partners with Meta After 36 years of licensing chip designs, Arm built its own chip: the Arm AGI CPU, designed for AI inference in data centers. Meta is the first customer and co-developer, planning to use it alongside hardware from Nvidia and AMD. The chip has up to 136 cores per CPU and claims double the performance per watt of x86 chips. This matters because Arm is now competing directly with many of its licensing partners. If you're tracking AI infrastructure costs, Arm is positioning this as an affordable option for companies that can't afford custom in-house processors. Other launch partners include OpenAI, Cerebras, and Cloudflare. Read more
OpenAI Shuts Down Sora Social App OpenAI announced it's shutting down Sora, the TikTok-like social app that launched six months ago. Despite impressive video and audio generation capabilities, user interest dropped sharply: downloads fell from 3.3 million in November to 1.1 million by February. The app's "cameos" feature (later renamed "characters" after a lawsuit) allowed users to create realistic deepfakes, leading to unchecked weirdness and moderation failures. Disney's $1 billion licensing deal, which would have let Sora generate Disney character videos, collapses with the shutdown. If you were planning to build on Sora's social features, that door just closed. The underlying Sora 2 video model remains available through other OpenAI products. Read more
Databricks Launches AI Security Product with Two Acquisitions Databricks announced Lakewatch, a new SIEM (Security Information and Event Management) product that uses AI agents powered by Anthropic's Claude to detect and investigate threats. The product is underpinned by two acquisitions: Antimatter (acquired last year, $12 million raised) and SiftD.ai (acquired this week). Both were small startups focused on secure agent deployment and interactive security notebooks. If you're using Databricks for data analytics, this extends the platform into security operations with AI-powered threat detection. Databricks indicated it's actively shopping for more acquisitions. Read more
Kleiner Perkins Raises $3.5B, Going All-In on AI Kleiner Perkins raised $3.5 billion across two funds ($1 billion early-stage, $2.5 billion growth), up from $2 billion two years ago. The firm has early stakes in Together AI, Harvey, OpenEvidence, Anthropic, and SpaceX. This signals continued VC appetite for AI investments despite the broader slowdown in exits. If you're building an AI startup, mega-raises from established firms like Kleiner Perkins, Thrive Capital ($10 billion), and Founders Fund ($6 billion) suggest capital is still flowing to proven AI bets. Read more
Secure Your AI Workflows with Dependency Cooldowns: If you use Python or JavaScript in production, configure your package manager to delay installing newly released packages by 2-7 days. This gives the security community time to spot supply chain attacks like the recent LiteLLM compromise. For pip, use --uploaded-prior-to with a cron job to update dates. For npm, set min-release-age in your config. For uv, use --exclude-newer with relative duration support.
Test Multi-Agent Workflows: If you already use Zapier AI agents, try building a two-agent workflow this week. Agent 1 handles data collection, Agent 2 processes and formats the output. Start simple: one agent extracts action items from emails, another creates tasks in your project manager. This pattern reduces errors compared to single-agent workflows because each agent has a narrower scope.
The AI control paradox is sharpening: Anthropic ships auto mode that decides when to ask permission, while security experts warn that giving AI "total control of your computer" will look foolish in retrospect. Meanwhile, shopping integrations stumble (OpenAI's Instant Checkout failed, Gap takes the Gemini bet), and AI social experiments like Sora collapse under their own creepiness. The infrastructure layer continues heating up with Arm's first chip and Databricks acquiring security startups. Watch for more supply chain attacks as AI tools proliferate, and for dependency cooldowns to become standard practice across package managers.