OpenAI voluntarily halted a massive RL run after its unreleased Astra model demonstrated autonomous zero-day exploitation and end-to-end novel cyberattack capabilities, triggering its prepa…

OpenAI's Astra model can reportedly identify and exploit zero-day vulnerabilities in real-world critical software autonomously, and execute end-to-end novel cyberattacks against hardened ta…

OpenAI's voluntary pause of a massive RL training run after Astra demonstrated autonomous zero-day exploitation and novel cyber attack capabilities shows frontier labs are operationalizing…

The voluntary White House testing framework for frontier models missed its August 1 deadline, remains classified (architected by NSA), and has not been shared with industry, creating uncert…

The White House's voluntary frontier model testing framework missed its August 1 deadline, remains undistributed even to briefed companies, and relies on classified NSA-architected benchmar…

Anthropic's IPO structure combines dual-class founder voting shares (ratio TBD) with a Long-Term Benefit Trust (including Ben Bernanke) that elects 4 of 7 directors — creating a novel gover…

Whisper differentiates in crowded voice AI by starting from product usage data (800k hours, 1.5x OpenAI's Whisper v3 training set) with opt-in privacy, targeting non-technical users through…

Whisper is building promptable, multimodal, duplex speech models (starting with Kanto) trained on 800K hours of opt-in user data—1.5x OpenAI's Whisper v3 dataset—to solve accent, noise, whi…

Whisper's 800k hours of opt-in user data (1.5x OpenAI's Whisper v3) and privacy-by-design architecture enables enterprise deployment, while its roadmap targets promptable speech models that…

Whisper's opt-in data sharing (5-10% opt-in rate) and on-device processing enable deployment in highly regulated environments like law firms and financial institutions, demonstrating a viab…