Welcome back! Congress gave OpenAI and Anthropic until August 24 to explain how their models got loose and hacked other companies. Elon's new bots sign into your accounts and keep working after you close your laptop, and a security firm needed less than a day to find a Zoom bug that hands your Mac to anyone in the call. Sam Altman, meanwhile, says ChatGPT is six months from watching your screen.
In today's Generative AI Newsletter:
OpenAI and Anthropic: What do House Democrats want by August 24?
SpaceXAI: What can Elon's new bots do while your laptop is closed?
Zoom: Why should you update before your next call?
Sam Altman: What will ChatGPT be watching six months from now?
House Democrats want Altman and Amodei under oath
House Democrats sent two letters Monday demanding OpenAI and Anthropic explain how their own models escaped containment and hacked other companies. Both set a deadline of August 24.
Rep. Greg Casar of Texas led it with Rep. Doris Matsui of California, and he's asking Speaker Mike Johnson to put both CEOs in a hearing chair.
Twenty-nine Democrats signed the letter to Sam Altman. It says a rogue model broke out of its security infrastructure for days without detection and hacked other firms from an internet connection it was never supposed to have. We covered that one when OpenAI admitted the model hit four more companies past Hugging Face.
The 22 who signed Anthropic's letter were blunter. Claude models got unauthorized internet access and hacked three companies on three separate occasions this year, and the lawmakers want the logs Anthropic still hasn't released, plus an answer on whether the company was negligent.
OpenAI says a review with outside advisers is underway and it'll publish the findings. Anthropic hadn't commented.
Special highlight from our network
Your team lives in Cursor and Claude. Your compliance tool doesn't.
AI-native teams are shipping faster than ever, but compliance still forces engineers out of their environment to check evidence, chase failing tests, and log remediation. That context-switch tax adds up.
Vanta now runs inside the tools your team already uses. This technical walkthrough shows how startups are using Vanta's MCP server, agentic workflows, and native plugins to keep compliance moving without leaving Claude, Cursor, or Codex.
You'll walk away knowing how to:
Connect Vanta into AI-native workflows through MCP and plugins
Surface failing tests, open risks, and audit gaps with a single prompt
Assign and track remediation without leaving your AI tool
Keep Vanta as your single source of truth, wherever your team builds
Can't make it live? Register anyway and get the recording.
Excel will never be the same
You already know which part of your month is the worst.
Cleaning the same messy export and rebuilding the same pivot before every report, while the work you're actually judged on gets whatever's left.
This course hands that loop to AI inside Excel.
3 live sessions with George Mount, a Microsoft Excel MVP whose courses have reached over 400,000 people, and you leave with a reporting process built on your own data that refreshes with one click.
Grok Bot signs into your apps and finishes the job
SpaceXAI opened an early beta of Grok Bot on Tuesday, AI teammates that each get their own computer in the cloud. You message one like a coworker, hand off a task and it signs into your real tools to do it.
Claude Cowork drives the apps on your own machine, so it stops when you do. Google's Gemini Spark has been running on cloud servers since May. What Grok Bot adds is the team part, several Bots working at once and passing tasks to each other in a group chat with you out of the middle.
There's nothing to configure first. You show a Bot the job once while you do it and it saves the steps, then runs them itself next time.
Inside SpaceXAI, a sales Bot updates the CRM from call notes and an ops Bot handles invoices sitting in Gmail. It also works on sites with no clean API, which is most of them.
Access comes with SuperGrok Heavy, Cursor Ultra and Cursor Teams Premium, and both the Mac download and the sales contact run through Cursor. Musk says the beta widens once the early bugs are fixed, with Grok 4.6 later this week.
Anyone in your Zoom call could have taken over your Mac
A security firm found a bug that let anyone sitting in your Zoom meeting take control of your computer with no click from you. Building the working exploit took under 24 hours and fewer than 20 prompts on publicly available AI models.
The flaw lives in Zoom's annotation feature, the drawing tools on a shared screen. Every client parses those messages automatically, so a crafted one corrupts the receiver's memory and runs code. It worked on Windows, macOS, iOS and Android. On a test Mac it popped Safari open on the victim's machine.
A Security found it on June 8 and had a working exploit the next day. Zoom shipped a client-side fix twelve days after the report and went public Tuesday. Zoom runs inside 70% of the Fortune 100.
Zoom's server-side filter can't see into end-to-end encrypted meetings, so anyone still on an old client is told to turn E2EE off until they're on 7.1.5 or 7.0.6.
Go update Zoom.
Altman says ChatGPT is six months from watching your screen
Sam Altman told a room of Silicon Valley interns we're six months out from a descendant of ChatGPT that watches your screen all day, sits in every meeting you're in, records every call and holds perfect context on your entire life.
He framed it as opt-in. You choose what to connect, texts, email, documents, Slack, and the assistant works from there.
His example was a startup CEO who can't possibly read every piece of customer feedback. An assistant with that much context writes the sales pitch off actual customer calls, or surfaces the complaint that keeps coming back.
The conversation was with Cory Levy at Internapalooza, the annual Bay Area intern gathering. One clip has cleared 430,000 views since Tuesday.
Everyone else in those meetings never agreed to any of it. Altman didn't get into that part.
Tool of the Day: Browser Use
Browser Use hands an AI agent a real browser and lets it click through websites for you. You describe the task in plain English and it navigates, fills forms, pulls data and reports back. The library is fully open source, so you can run the whole thing on your own machine with your own Anthropic, OpenAI or Google key and pay nothing beyond your model provider. The hosted version adds stealth browsers at $0.02 an hour with proxies in 195+ countries for sites that block bots.
Try this yourself:
Run pip install uv, then create and activate a virtual environment with uv venv --python 3.12.
Install the library with uv pip install browser-use, then run uvx browser-use install to pull down a Chromium.
Drop your Anthropic, OpenAI or Google API key into a .env file, or use a Browser Use key for their own model.
Copy the starter script from the docs and swap the task line for something you actually do, like finding the cheapest flight from Berlin to Lisbon next Friday.
Who it's for: anyone repeating the same 20 clicks on the same site every week. Skip the install and the free cloud plan starts at 10 agent tasks a month.
Everything else you shouldn't miss
Sam Altman's right hand is walking out: Brad Lightcap joined OpenAI in 2018 and ran operations until April, and he told staff Tuesday he's leaving in a few weeks to start something new.
Researchers pulled the hidden thinking out of Claude, ChatGPT and Gemini: They replayed a top model's encrypted reasoning block into its weakest sibling and jailbroke that one into transcribing the whole thing, and all three labs have since shut it down.
Gemini hit a billion people: Google says the app crossed 1 billion monthly users, the fastest growth of anything it has ever shipped, with 63% of them talking to it out loud.
Nvidia dropped a free model that runs on one GPU: Nemotron 3.5 Lightning is 30B parameters with 3B active per token, free for commercial use, and Nvidia claims up to 4x faster output than open models its size.
Spotify will tag AI artists and stop pushing them: AI Persona badges land in mid-September and those profiles drop out of editorial and algorithmic recommendations unless you already follow them.
Learn more about AI from the experts building it
📸 Follow us on Instagram for fast, visual AI updates in 30 seconds.
📺 Watch us on YouTube to hear insights directly from leading AI voices, builders, and innovators.
🐦 Follow us on X for breaking AI news and real-time industry updates.
🧠 Learn how to build your next AI application with practical resources and expert guidance.
