Apple opens Siri AI to every app, with SKILL.md style skills for multi-step tasks
•Season 3•Episode 35
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
0:00
|
6:12
Top News: Apple opens Siri AI to every app, with SKILL.md style skills for multi-step tasks
Plus: CrowdStrike buys two AI security startups for $495 million and launches an AI Kill Switch for agents
Top AI Tools:
Wanaka Studio (https://wanaka.app/) is an AI game engine where an agent builds a 3D game for you, then lets you edit it in a synced visual editor with auto rigging, multiplayer, and one click publishing, no code needed.
Floot Payments (https://floot.com/) lets apps you build inside ChatGPT, Claude, or Cursor take real money, setting up subscriptions, one time payments, coupons, refunds, and shipping on Stripe without API keys or webhooks.
Prelulu (https://prelulu.ai/) brings interactive stories with AI characters you can text or video chat with, using real time avatars, and lets you create your own characters.
Business AI Tip of the Day
Give Every Client Agent a Kill Switch Before It Goes Live
SPEAKER_00
Welcome back to Sumo Growth AI Insights. I'm Darius and I'm walking you through this week's AI news the way an operator actually uses it, not as a pile of headlines. Lead story. Apple opens Siri AI to every app, with skill files for multi-step tasks. Apple opened a Siri AI developer preview to every developer on Tuesday, starting with the iOS 27.3, iPad OS, Mac OS, and Vision OS 27.3 betas. Any app can now hand its content and actions to Siri AI through App Intents, so people can just ask in their own words to check a delivery, book a reservation, or place an order, and Siri uses personal context and on-screen awareness to carry it out. For bigger requests, Apple added app skills that teach Siri how to chain several actions together, and developers have noticed they're written as skill.md files, a narrower version of the same agent skills format coding agents already use. Testing is limited to development builds and internal test flight for now, with customers getting it in a few months and app store submissions opening in the spring. For agencies, the phone in every client's pocket is turning into an agent, and the businesses whose apps and booking flows are ready for it will get the voice requests first. Plus story CrowdStrike buys two AI security startups for $495 million and launches an AI kill switch for agents. CrowdStrike announced deals for Fiddler AI and DreadNode at its Falcon Europe event, with the price reported at about $400 million and $95 million. Fiddler watches agents after they go live, tracing their decisions and flagging drift and made-up answers, while Dreadnote runs more than 100 attack methods, like jailbreaks and prompt injection, against agents before they ship. CrowdStrike also launched an AI kill switch in Falcon Guardian that cuts off an agent's access to its models through an AI gateway, with endpoint shutdown coming soon, plus hooks that enforce policy inside Claude Code, Codex, GitHub Copilot, and Cursor. CEO George Kurtz put it simply. The more certain a company is that it can stop any agent, the more agents it can put to work. For agencies selling automation, a clear way to pause or shut off every client agent is quickly becoming part of the pitch. Three more headlines. First, Tencent launches HY4, an open model with vision and computer use. Tencent released HY4 on Wednesday, adding image understanding and computer use for the first time, trained on professional software like Blender and CAD. Weights and system cards are promised in the coming days. API pricing stays around 90 cents per million input tokens and $2.70 per million output tokens, and it's coming to OpenRouter. Second, ChatGPT hits 1.2 billion weekly users and says its search is about a fifth of Google's. ChatGPT consumer lead Adam Fry said weekly users grew from 1 billion to 1.2 billion in two months, and that ChatGPT's searchlight queries now equal roughly 20% of Google's volume by OpenAI's own count. He also said 250 million people use voice each week, and that chat and work are merging into one experience. Third, Anthropic aims to double its computing power to more than 7 Gigawatts next year. Anthropic expects more than 3.5 gigawatts of compute by the end of this year and wants to roughly double it in 2027. Its IPO filing lists $895.4 billion in compute and data center commitments through 2045, including deals with Google, Amazon, Microsoft, and SpaceX. Tools of the Day. Wanaka Studio at wanaka.app is an AI game engine where an agent builds a 3D game for you, then lets you edit it in a synced visual editor with auto rigging, multiplayer, and one-click publishing. No code needed. Flute Payments at Flute.com lets apps you build inside ChatGPT, Claude, or Cursor take real money, setting up subscriptions, one-time payments, coupons, refunds, and shipping on Stripe without API keys or webhooks. Pralulu at Pralulu.ai brings interactive stories with AI characters you can text or video chat with using real-time avatars and lets you create your own characters. Business AI tip of the day. CrowdStrike just made the case that being able to stop an agent is what lets companies run more of them. Your clients feel the same way. An AI receptionist or follow-up agent that can't be paused in seconds is a risk, and one that can is an easy yes. So add a simple kill switch to one client agent this week. Create one on and off field, like a custom value or tag in sumo AI CRM or Go High Level that every agent workflow checks before it acts. Point all agent model calls through one API key or gateway per client so you can revoke access in one place. Set hard limits, like a maximum number of messages per contact per day, and a spending cap on tokens. Write a short pause playbook that says who can flip the switch, how the client gets told, and how calls fall back to a human. And test it once a month by switching the agent off and confirming nothing goes out. Clients sign off on automation faster when they know exactly how to stop it. Show them the switch during onboarding, not after something goes wrong. Pro tip log every time the switch is flipped and why. Those notes become your best guide for tightening the agent's rules. That's your sumo growth AI Insights briefing. Build that kill switch this week, and I'll catch you on the next issue.