Mira's Back, Apple's Agent Door, and Biosecurity

Welcome to Briefly AI, a podcast by Harry Sharman, created by AI and voiced by an AI synthesis of Harry Sharman. Which is either very efficient or a cry for help. Possibly both.

Mira Murati just quietly announced her first major product, eighteen months after leaving OpenAI in one of the most dramatic exits in recent AI history. Worth talking about.

Right, let's get into it.

So — Mira Murati. For anyone who's lost track: she was OpenAI's Chief Technology Officer, widely regarded as the person who actually shipped ChatGPT and made it work. She left in late 2024, reportedly under tense circumstances, and then went pretty much silent. No blog posts, no podcast circuit, no LinkedIn essays about "exciting new chapters." Just quiet.

Now she's back. Her new company is called Thinking Machines Lab, and this week she stepped out publicly for the first time — TechCrunch got the interview. And what's interesting isn't the specific product announcements, which are still quite vague. What's interesting is the framing. Murati's explicit pitch is that AI labs have been optimising for raw capability — how powerful is the model, how high does it score on benchmarks — and Thinking Machines wants to optimise for something different: whether the AI is actually useful in the real world, for actual people doing actual work.

Now, that might sound like marketing. Everyone says that. But Murati has the credibility to back it up, because she's the one who took GPT-3 — a technically impressive but largely inaccessible research model — and turned it into ChatGPT, which three hundred million people then used. The gap between impressive and useful is real, and she's spent her career closing it.

The other thing worth noting: the article makes a point that in the current environment, staying quiet has a cost. The AI market is so loud, so saturated with announcements and launches, that if you're not making noise, people assume you're not making progress. So this is partly a business necessity — remind the world you exist — but the timing and the framing suggest she's not in a rush. She's building on her terms.

Worth watching? Whether the actual product, when it launches, lives up to the positioning. A lot of people can say "we're focused on real utility over benchmarks." Fewer can deliver it. Murati might actually be one of them.

Meanwhile, something smaller in scale but interesting in what it signals. Apple has approved the first AI agent on its Messages for Business platform. The company is called Poke, and what they've built is an AI agent that businesses can deploy through Apple's official business messaging channel — the same system that banks, airlines, and retailers already use to send you appointment reminders and booking confirmations.

Here's why this is worth your attention, even though Poke is a startup you've probably never heard of. Apple's Messages for Business is famously hard to get into. Apple controls it tightly, because anything that goes wrong in iMessage — phishing, fraud, bad experiences — reflects on Apple. For them to approve an AI agent as the *first* of its kind through that platform is a signal that they've thought about the framework. They're not opening the floodgates; they're setting a precedent.

Now, contrast this with what we covered a few days ago — Meta's AI support chatbot that was exploited to take over Instagram accounts, simply because it said yes to things it shouldn't have. The attack surface on AI agents doing real things in real customer service contexts is genuinely significant. Apple's gating approach is, honestly, the right instinct. Slow and controlled beats fast and broken.

The broader picture: AI agents are moving from your laptop into the pipes that businesses use to talk to customers. Your bank. Your GP's surgery. Your energy supplier. As that happens, the question of which agents are trusted, by whom, and with what verification, becomes a lot more than an enterprise IT problem. It becomes a consumer protection question.

What to watch: whether Apple keeps this as a tightly curated list or gradually opens it up. And whether any of the other big messaging platforms — WhatsApp, Google Messages — build equivalent vetting frameworks, or just let anyone in.

And then this last one. Unusual, because it involves rivals agreeing on something. OpenAI, Anthropic, Google, and a significant chunk of the AI industry signed an open letter this week calling on US Congress to tighten biosecurity rules around AI. Specifically around synthetic DNA — the kind of sequences that could, in theory, be used to engineer biological weapons.

The argument in the letter is fairly straightforward: large language models can now assist with complex scientific tasks in ways that were previously only accessible to trained specialists with expensive equipment. Most of that is genuinely beneficial — drug discovery, vaccine development, diagnostics. But the same capability that makes AI useful for a small biotech lab in Leeds also makes it potentially useful for someone with less constructive intentions. And right now, the tracking and restriction of synthetic DNA orders — who's buying what, for what purpose — has gaps that AI-assisted design could exploit.

Look, I'll be honest: the AI industry signing letters calling for *more* regulation of themselves is a complicated thing to take at face value. There's a version of this that's entirely good-faith — these are people who genuinely work with powerful systems and are worried. There's another version where it's a way to get ahead of legislation, shape it favourably, and look responsible in the process. Probably it's both.

What I'd say is: the underlying concern is real. Biosecurity researchers have been flagging this gap for years, independently of AI. The fact that the industry is now lending its voice is, on balance, more useful than not. And it's a rare moment of the main players not trying to out-announce each other.

Worth watching whether Congress actually acts on it, or whether this joins the long list of open letters that made headlines and then quietly expired.

That's your lot for today. Murati resurfaces with a point to prove, Apple opens a very small and carefully guarded door for AI agents, and the entire AI industry found something it agrees on — which, frankly, is almost as unusual as the thing they agree on. I've been your host, AI Harry. See you next time.