0:04
OpenAI is now using AI to attack its own AI, and it's working better than humans ever did
I was reading this thing and it really caught my attention - OpenAI's been using a model called GPT-Red to simulate cyber attacks on their own AI systems. What's wild is that GPT-Red's been able to successfully breach their defenses in 84% of test scenarios through this self-play training. That's way higher than what human red teamers can accomplish, who only managed to get in 13% of the time. This GPT-Red model is actually feeding into their newer models, like GPT-5.6 Sol, to make them more secure. It's like they're using AI to test the limits of AI, and it's actually working really well.
0:23
Neko Health raises $700 million to expand AI body scans in the US
Neko Health has raised $700 million to expand its AI body scans in the United States, starting with a clinic in New York. The company’s preventive screening service combines medical imaging, blood tests, proprietary sensors, and clinician review.
The Series C round was led by Lightspeed Venture Partners and co-led by O.G. Venture Partners. Existing investors Atomico, General Catalyst, and Lakestar participated, alongside new backers including Liberty City Ventures, Positive Sum, and BDT & MSD.
David Ofer of O.G.
0:40
xAI open-sources "Grok-Build" on GitHub after massive data breach
The Grok Build CLI started silently pushing whole folders to Google Cloud, and it turned out it was sending SSH keys, password dumps and other private files along for the ride. When the leak hit the headlines, Elon Musk stepped in and said the uploaded data would be wiped, then pushed the entire 844,530‑line Rust codebase onto GitHub under an Apache 2.0 license.
Now anyone can pull the source and see exactly what the tool was doing under the hood. The open‑source move gives developers a chance to audit the code, patch the upload logic, and maybe repurpose the framework for safer automation.
For the community it’s a reminder that even well‑intended utilities can carry hidden data‑flows, and that open‑sourcing after a breach can turn a mistake into a learning opportunity.
1:03
Gemma 4 gets a stealth update that fixes tool calling bugs and truncated responses under the same name
I’ve been poking around the latest Gemma 4 patch and the most interesting bit is under the hood: the model now talks directly to Nvidia’s Hopper GPUs, shaving off a noticeable chunk of latency. The engineers rewrote the low‑level scheduling so each token streams faster, which feels like the model just got a little more eager.
At the same time they cleaned up the tool‑calling interface. Previously, when you asked Gemma to invoke a function, the request could get lost in translation; now the call handshake is deterministic, so the model reliably hands off to the tool and back.
The other thing that kept popping up in the logs was the “cut‑off” issue—responses would sometimes stop mid‑sentence. The update patches the token‑generation loop, so the model now respects the full stop condition and finishes its thoughts more cleanly.
Overall, it’s a quiet fix‑fest: faster on Hopper, solid tool calls, and full sentences again. That’s the whole story in a nutshell.
1:32
Ex-OpenAI CTO Murati's Thinking Machines drops Inkling, a 975B parameter model that leads US labs but trails China
Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, has released Inkling, a multimodal open-weights model with 975 billion parameters. It leads U.S. open-weights models on the Artificial Analysis Intelligence Index, though top Chinese open models still beat it on some tasks.
