AI Signal Brief: Real-World Threats, Breakthroughs, and the Shifting Landscape

Google’s Threat Intelligence Group just detected and stopped the first known AI-assisted zero-day exploit attempt. In our last episode, we talked about cutting through the hype to find the signal. Today, that signal is a siren. The theoretical threat of AI discovering and weaponizing unknown software flaws is now a documented, real-world attack. That’s the headline that changes the security landscape. But the ground is shifting everywhere. Thinking Machines Lab, the new company from former OpenAI CTO Mira Murati, just released a multimodal model that can interact in real time.

We’re talking two hundred millisecond latency—fast enough to interrupt you mid-sentence based on a visual cue. It’s a genuine breakthrough in interaction, even if the company itself is on shaky ground after a few co-founder departures. Meanwhile, the money is still moving. Isomorphic Labs, the drug discovery spin-out from DeepMind, just pulled in two-point-one billion dollars in a Series B round. That’s not just a vote of confidence; it’s a declaration that AI-native drug design is its own scalable industry now. And on the other end of the cost spectrum, a new company called Perceptron released a video analysis model, Mk1, that claims state-of-the-art performance for eighty to ninety percent less than OpenAI or Google.

Their price is just fifteen cents per million input tokens. That’s how you commoditize a market overnight. Finally, we’re seeing the strategic map being redrawn. OpenAI just launched GPT-5.5-Cyber for vetted teams in the European Union, getting a critical first-mover advantage there while Anthropic’s competitor has no announced EU rollout. And Google is hiring hundreds of “forward deployed engineers” to embed AI directly with customers, which tells you everything you need to know about the complexity of moving from a simple API to a deeply integrated workflow. Let’s go back to that zero-day exploit.

For years, the debate around AI safety has included the risk of models autonomously finding novel software vulnerabilities. The concern was that a sufficiently powerful AI could discover flaws that no human had ever seen, and then write the code to exploit them. It was a plausible, but theoretical, danger. Not anymore. Google’s report confirms it. They stopped the attack before it executed, but the key phrase in their statement is this: AI-assisted zero-day discovery has "moved from theoretical concern to documented real-world attack.” This wasn’t a red-team exercise or a research paper.

This was a live weapon, built by an AI, aimed at a real target. The offensive and defensive cybersecurity arms race just entered a new, much faster phase. Now let’s talk about that new model from Thinking Machines Lab. TML-Interaction-Small. The name is terrible. The technology is not. The benchmark to watch is turn-taking latency. How long does it take for the AI to respond after you stop talking? Google’s latest Gemini is at zero-point-five-seven seconds. OpenAI’s real-time model is a sluggish one-point-one-eight seconds. This new TML model clocked in at zero-point-four-zero seconds.

But that number hides the real story. The model processes everything—audio, video, text—in two hundred millisecond chunks, simultaneously. This means it doesn't wait for you to finish your turn. It can see you frown, hear the hesitation in your voice, and interrupt you with a clarifying question before you’ve even finished your sentence. This isn't a better chatbot. This is a fundamentally different mode of interaction. It moves the AI from a tool you address to a presence you collaborate with. Of course, the breakthrough comes from a company that’s had a rocky start, which poses the real question: can a brilliant piece of tech survive an unstable home?

For a long time, the most advanced AI lived in a simulated world. It played games, it wrote stories, it answered hypothetical questions. That world is gone. Today, AI is discovering new medicines with billions in backing, and it’s discovering new weapons being deployed in the wild. The models are no longer just processing the world; they are becoming active agents within it. The training wheels are off. This is the real thing.