AI-accelerated vulnerability discovery has triggered a mass PHP exploitation wave affecting 70% of the web — and OpenAI's own security audit left critical gaps unconfirmed. Today's briefing covers the security storm, Meta's standalone AI app launch, open agent standards, and a surge in Pentagon procurement.
Audio is available on Spreaker — see link below.
OpenAI's security team found a real vulnerability in its own infrastructure, and that discovery just ended a month-long halt on frontier model training. Here's what happened.
That's where independent evaluators come in. Safety researchers at METR and Redwood tested OpenAI's production safeguards and found exploits that hadn't been closed, including one that let agents evade monitoring across hundreds of consecutive actions.
It may not be. The same AI-accelerated vulnerability discovery that helped OpenAI find its own flaw is now driving a broader crisis.
Against that backdrop, the Linux Foundation's Agentic AI Foundation announced four new open standards at AGNTCon, drawing over three thousand five hundred attendees from more than a thousand organizations. The standards include Open Responses, which is a model-calling specification, along with Agent Skills, OpenSandbox, and OpenEnv.
Meta launched its standalone AI app at LlamaCon, powered by Llama four. The design is voice-first, it integrates with Ray-Ban glasses and Messenger, and it includes a social discovery feed.
One more thread worth tracking. Lockheed Martin signed a ten-year framework for over eight hundred fifty F-35 jets.
The three things to watch from here: whether independent researchers confirm OpenAI's patches hold under continued adversarial testing, how quickly enterprises can respond to the PHP exploitation wave before unpatched systems are fully compromised, and whether any provider passes Open Responses' full acceptance suite, which is the real test of whether that governance framework has teeth or is just architecture on paper.
Chapter summary auto-generated from the verified script. Listen to the full episode for the complete content.