In partnership with
FROM OUR PARTNER
Stop typing AI prompts. Start talking.
You think 4x faster than you type. So why are you typing prompts? Wispr Flow turns your voice into ready-to-paste text inside any AI tool. Speak naturally, tangents and all, and Flow cleans it up. Available on Mac, Windows, iPhone, and Android.
TODAY AI BRIEF
3 things that happened while you were busy
1 Google launches an AI that finds and patches security flaws on its own.
Google just released Gemini 4 Argon in limited access, aimed at complex engineering, enterprise and cybersecurity work. It hunts down a serious security flaw, confirms the flaw is real, and writes the fix without a human steering, and it leads the DeepSWE coding test at 77.9%. See how Google keeps made-up answers to a minimum.
2 Two coding startups are feuding in public over a board advisor.
On Wednesday, Factory CEO Matan Grinberg said he had fired board advisor Chris Degnan for allegedly leaking secrets to its biggest rival, Cognition. Two hours later, Degnan announced he had joined Cognition as revenue chief and had resigned on his own, while Cognition's CEO denied receiving any Factory information. See the full timeline.
3 SpaceXAI ships shared AI teammates that you can tag in Slack.
SpaceXAI launched Team Bots on its Teams and Enterprise plans, with ready-made bots for project management, testing and fact-checking. You can tag one in Slack, and it hands the actual coding to Cursor agents while tracking the whole workflow. Watch how one senior engineer uses it every day.
Bonus reads: Google also added Skills, which turn a task you repeat into a reusable template. Watch Skills in action, plus try it yourself. Developers can also watch eight practical ways to use Jev, a fast decision model for quick, structured calls.
AI FORWORD
Your AI agent is only as safe as the rules someone wrote for it
The problem. Give an AI agent too much freedom and it can wipe files, run up a huge bill, or call a job finished when it is not. Most ready-made agents bury their limits in default settings your team never picked. ML researcher Elvis Saravia's new walkthrough starts from exactly that worry.
The fix. The answer is a harness, a set of ground rules wrapped around the agent. Saravia builds one with the Pi SDK and Jev, so that before any tool runs, a small check decides whether to allow the call, block it, or ask a person.
The bill. Jev also routes simple requests to a cheaper model and saves the stronger one for hard work, which keeps costs down. OpenAI's new cookbook makes the money case: cached input tokens, the chunks of text AI bills you for, can be up to 90% cheaper when agents reuse shared instructions.
The paper trail. A final pass compares the agent's reply with the files it read, and a retry limit stops it from looping forever. Every decision gets logged along with the numbers behind it, and plain code blocks obvious hazards like touching files outside the project. You can poke at all of it in a live sandbox that includes a delete tool.
The reframe. The interesting question is no longer how smart an agent is, because its limits can now be read in code instead of hidden in settings. That matters more as tools like Open Dot, a Mac app where each agent gets its own logged in browser and can hand work to other agents, catch on. Before you give any agent your inbox or your store, ask who wrote its rules.
PRESENTED BY :
These 7 Stocks Are Built to Outlast the Market
Some stocks are built for a quarter… others for a lifetime.
Our 7 Stocks to Buy and Hold Forever report reveals companies with the strength to deliver year after year - through recessions, rate hikes, and even the next crash.
One is a tech leader with a 15% payout ratio - leaving decades of room for dividend growth.
Another is a utility that’s paid every quarter for 96 years straight.
And that’s not all - we’ve included 5 more companies that treat payouts as high priority.
These are the stocks that anchor portfolios and keep paying.
You can download this report for free as of today, but it won’t be free forever.
This is your chance to see all 7 names and tickers - from a consumer staples powerhouse with 20 years of outperformance to a healthcare leader with 61 years of payout hikes.
SOCIAL SNAPSHOT
What people are actually watching and sharing
Napoleon's secret letter. A coded letter to one of Napoleon's generals stayed unreadable for 217 years. GPT-6 Astra produced the full decoded version in six hours (2.5M views).
Apple's next gadget. Apple is releasing its next big home hardware item on October 13, and one design detail throws straight back to Steve Jobs' 2002 iMac (1.5M views).
Sonnet 5.5 verdict. A senior developer tested Claude Sonnet 5.5 for a full day and says the "cheaper Opus" label is wrong. His notes show exactly where it falls apart (1,600 likes).
Game dev with a catch. An Anthropic engineer's case against handing game development to AI is going viral, backed by his own brawler prototype (3,500 likes).
Rules from the 90s. An ex Microsoft and Google engineer turned his '90s principles into hard rules for Claude Code and says his cost per feature dropped from $640 to $1 (1,000 likes).
App Store survival guide. A senior dev shared a launch checklist for anyone shipping an app to the App Store or Play Store, so no hidden requirement catches you late (2,200 likes).
PROMPT STATION
Write the ground rules for any AI agent in seconds
Today's harness story boils down to one idea: decide what an agent may do before it does anything. This prompt works in Claude, ChatGPT or Gemini and turns that idea into a plain English rulebook for any agent you plan to use at work. Run it before you hand an agent access to your email, files or store.
You are a careful operations advisor. I am about to give an AI agent access to [TOOLS OR APPS THE AGENT CAN USE]. Its job is [WHAT THE AGENT WILL DO].
Write a one page rulebook for this agent in plain English. Sort every action it might take into three groups: ALLOW (safe to do on its own), ASK ME FIRST (needs my approval), and BLOCK (never allowed).
For each group, give five specific examples tied to my tools. Then add a spending limit of [MAXIMUM MONTHLY BUDGET] with a warning at 80 percent, a rule for what the agent must do when it is unsure, and a simple log format so I can review every decision it made.
Finish by listing the three riskiest things this agent could do and how the rulebook stops each one.