Welcome back! Google says the Gemini app has now passed 1 billion monthly users, making it one of the company’s fastest-growing products ever. People are also talking to it a lot: 63% of Gemini users use voice, and the app now generates more than 150 million images a day.
At this point, the AI adoption curve is less “early experiment” and more “everyone is casually asking a chatbot to make them a dragon LinkedIn headshot before breakfast.”
Is "Our AI Broke Containment" the New AI Benchmark?
Lately, it feels like every major AI company is competing on the weirdest possible metric: whose model can commit more cyber crimes.
Here's the running scoreboard:
July 21: OpenAI disclosed that two of its models, including GPT-5.6 Sol, exploited a zero-day to escape their sandbox and hack Hugging Face (we covered that one in depth a couple weeks back).
July 30: Anthropic followed with an internal review of 141,006 evaluation runs and disclosed that three of its Claude models—Opus 4.7, Mythos 5, and an internal research model—breached real production systems at three different organizations during capture-the-flag tests.
Aug. 5: Meta joined the party. Its recently released Muse Spark 1.1 model hacked an unnamed third-party service during cybersecurity testing.
There's a wrinkle worth flagging though. Anthropic's and Meta's incidents both trace back to the same source—a small (~35-person) AI evaluation firm called Irregular, which had a misconfiguration that accidentally gave the models internet access during what were supposed to be sealed tests. So while OpenAI's case was a real sandbox escape (the model exploited an unknown vulnerability to break out), the Anthropic and Meta cases were more "the door was left open and the model walked through it."
My POV: The actual cybersecurity implications here are serious. Real production systems at real companies got accessed. Some of those companies didn't even know it had happened until Anthropic called them.
But I also can't shake the feeling that Meta may have been at least partly using this disclosure to signal how capable its models are. As the company fights to get back into the frontier AI conversation, "our model is smart enough to break containment too" carries a very specific marketing energy.
We probably shouldn't be celebrating containment breakouts as a sign of intelligence. But that's where we are. I just hope we're not playing with fire.
What do you think: healthy dose of transparency, or the first sign of a much bigger problem? Hit reply with your thoughts.
— Matt
Grok Is Now an AI Teammate
Via HYPEBEAST
SpaceXAI launched Grok Bot, an always-on AI agent service that lets users assign work to AI “teammates.”
What it does:
Grok Bot gives users AI agents that can work independently inside their own cloud-based computer environment. The bots can sign into apps, websites, and tools, complete tasks, and report back when something is finished or needs approval.
Users can message the bots like colleagues from desktop or mobile without building workflows in advance. The bots can also message each other, coordinate in group chats, share context, and assign ownership when projects overlap.
Multiple bots can run in parallel, and one bot can even manage others working on specialized tasks.
Grok Bots can learn existing workflows, preserve a user’s personal voice, pick work back up from old chats, follow up on dropped conversations, and become more proactive over time.
The idea is not just “answer this question,” but “remember how I work and start doing the annoying stuff before I ask.”
The competition: Grok Bot joins a growing field of workplace agents from major AI labs, including OpenAI’s ChatGPT Work, Anthropic’s Claude Cowork, and Microsoft’s Copilot Tasks.
The bigger picture: Everyone is racing toward the same interface: not a chatbot you prompt, but an AI coworker you assign. The next wave of assistants will not just draft emails or summarize documents. They’ll log into tools, coordinate projects, remember workflows, and act on your behalf. That could make work feel much lighter—or make everyone realize they now have a fleet of extremely eager interns with browser access.
The ChatGPT Dog Cancer Story Becomes a Startup
The viral story about an entrepreneur using AI tools to design a personalized cancer vaccine for his dog has now turned into a startup.
Backstory: Earlier this year, Australian tech entrepreneur Paul Conyngham drew attention after saying he used ChatGPT, Grok, and computational genomics to design a personalized mRNA cancer vaccine for his dog, Rosie. He credited the treatment with shrinking several tumors and adding years to her life.
Messy science: The story was always more complicated than the viral version suggested. The exact role AI played was unclear, and researchers warned that the claims gave too much credit to the tools while minimizing the human expertise involved. Rosie also received another therapy alongside the vaccine, making it impossible to say whether the mRNA treatment itself caused the improvement.
The startup: Conyngham is now launching Gamgee, a Y Combinator-backed startup focused on personalized mRNA cancer vaccines for dogs. Gamgee says veterinarians will be able to submit cases with less than 20 minutes of work, then receive a personalized vaccine in around four weeks. The company says it wants to use AI and genetics to build treatments across diseases and species, eventually including humans.
The bigger picture: This may sound like peak startup theater: AI, dogs, mRNA, Y Combinator, and an anime launch video all in one story. But underneath the hype is a real direction worth watching. Personalized medicine is one of the places AI could eventually matter most, and even messy early attempts can push research, trials, and public attention forward. The question is whether Gamgee can turn a viral anecdote into actual science.
Build a Team of AI Teammates Your Whole Team Can Use. Together.
Stop juggling prompts and start running a real crew of always-on agents. With Hyperagent, you have a specialized team to get work done while you literally sleep.
Proactive with compounding skills, Hyperagent can:
Spin up the right agent for every part of the job.
Give each agent tools and its own login across Slack, Gmail, Docs, Airtable, and more.
Allow agents to hand off work to each other. Output from one becomes input for another.
Let your entire team use the same agents directly in Slack or wherever they need.
Ready to assemble your AI team? Get $100 in bonus credits when you sign up for a paid plan.
Multiplayer AI Coding Sessions
Via mpai
mpai is an open-source terminal tool that lets a named teammate join an existing Codex or Claude Code session from their own terminal. The host shares a specific session with one command, and the guest joins through a private invite over a Tailscale network.
How you can use it
Let coworkers join active Codex or Claude Code sessions
Collaborate without screen sharing or pasting transcripts
Attribute each participant’s prompts by name
Keep session control with the host
Share agent work privately on a per-session basis
Pricing: Free
Run Local AI Models on Your Mac
Via Nativ
Nativ is a native macOS workspace for running AI models locally with no accounts, subscriptions, or cloud dependency. It includes an MLX-VLM inference server, a curated model library, and a SwiftUI chat interface for language, vision, code, and embedding models.
How you can use it
Run AI models locally on Apple Silicon
Use language, vision, code, and embedding models in one workspace
Connect Claude Code or Codex to locally served models
Use it as an OpenAI- or Anthropic-compatible local inference server
Inspect the open-source stack on GitHub
Pricing: Free
Jobs, announcements, and big ideas
OpenAI ships GPT-5.6 Sol Ultrafast mode, streaming 750 tokens per second on Cerebras chips—14x faster.
Google debuts Gemini 3.7 Flash, delivering major coding gains at half the price of its predecessor.
Anthropic upgrades Claude Tag with channel-wide context, making it 30% better at knowing when to respond.
DeepSeek releases V4-Pro, pairing agent upgrades with flexible reasoning and revamped API pricing.
Alibaba unveils Wan3.0, generating AI video from any input in 30 seconds on Cloud Model Studio.
Anthropic brings Claude Cowork to Chrome's side panel, syncing sessions across desktop and mobile.
The explosion 💥 of the AI agent race, and the coolest AI news from this past week in my video—check it out!
That’s a wrap! See you next week for more.
