OpenAI safety writer quits, calling its culture broken
David Robinson says trial-and-error deployment guarantees larger failures, plus India's removal of Bitchat and Google's new Gemini model limits.
On this page
David Robinson, who says he led safety reports for OpenAI’s major launches over three and a half years, has resigned. He wrote in The Atlantic that trial-and-error deployment guarantees periodic failures that will grow with more capable systems. He cited the recent breach of Hugging Face systems by OpenAI agents and continued reports of rogue behavior.
OpenAI spokesperson Drew Pusateri said the company pauses training or holds back models when needed. The company says it is expanding third-party evaluation and real-time monitoring. Robinson’s account of culture and risk is his own assessment, and he acknowledges he hired a public-relations firm to speak out.
Featured source: TechCrunch , The Verge .
Other Stories
Bitchat disappears from Indian app stores after government order
Jack Dorsey said India’s Ministry of Electronics and IT ordered Apple to remove Bitchat in India under Section 69A of the IT Act. Apple says the app remains available outside India, while TestFlight access in India is also blocked. Google Play showed the app as unavailable in India when TechCrunch checked, though it could not confirm whether Google received the same federal direction.
Bitchat, launched in July 2025, sends encrypted messages over Bluetooth mesh without cellular, internet, or central servers. India previously ordered GitHub to remove related repositories in July 2026 over interception concerns. The restriction coincides with youth protests and a 12-hour internet suspension near New Delhi’s Jantar Mantar. Neither the order nor the cited unlawful content has been made public.
Filed from: TechCrunch , Internet Freedom Foundation .
Bug-hunting model found a flaw that came under attack within a day
Horizon3 researcher Zach Hanley used Anthropic’s Mythos model through Project Glasswing to find an authentication bypass in Rejetto HTTP File Server. The flaw, tracked as CVE-2026-61500, can lead to full admin access and remote code execution and is fixed in version 3.2.1 and later. VulnCheck researcher Patrick Garrity said exploitation began the next evening, first from one China-hosted address against hosts in the United States and Japan.
The report links Mythos and Glasswing to 286 CVEs, with this the second seen exploited in the wild. Attacker attribution remains uncertain because actors often route traffic through compromised proxies. The technical account of the weakness comes from Horizon3’s write-up as reported, not from independent testing here.
Filed from: The Register , Rejetto .
Google limits free and Plus users to weaker Gemini models
Google is changing which Gemini models individual users can access by payment tier. Free users will be restricted to the Flash-Lite (3.5) model and lose Flash (3.6) and Pro (3.1). AI Plus subscribers at $4.99 per month keep Flash-Lite and Flash but lose Pro (3.1), while Pro and Ultra tiers keep their models.
Google says the free-tier change takes effect October 9, with Plus users to be notified by email. The move was described from an updated support document via 9to5Google rather than independent testing of tier enforcement here. Exact query limits remain unclear because Google says they depend on prompt complexity, features used, and chat length.
Filed from: TechRadar , Google support document .
Valar Atomics proposes 456 reactors for a Utah data-center campus
Valar Atomics has proposed placing 456 nuclear reactors on more than 9,000 acres of federal land near Price, Utah. The plan adds about 650 acres of state land. The staged modular build, called Project Beehive, would supply a computing campus with 9.6 gigawatts. At continuous output, that is about 2.4 times Utah’s roughly 4GW average generation. It would be enough for roughly 7 million homes, with none of the power going to homes.
The figures as reported contain an unresolved discrepancy, since 456 units at 25MW each would total 11.4GW rather than 9.6GW. The item remains a single-source proposal report with no approval or construction timeline disclosed, so it should be treated as proposed rather than confirmed.
Filed from: TechRadar , Valar Atomics .
Old interface laws resurface after a NeoVim data-loss complaint
Computer scientist David Chisnall says trying NeoVim overwrote the persistent-undo file from his long-used Vim setup. He says the loss destroyed more than a decade of undo history. Designer Marcin Wichary uses the episode to revisit Jef Raskin’s three laws from the 2000 book The Humane Interface.
The laws hold that software must not harm work, must not waste time, and must respect human needs and frailties. This is commentary rather than a new finding, and the account comes from Chisnall’s own post as quoted secondhand. There was no independent verification of the NeoVim behavior here, and later updates acknowledge a warning was shown.
Filed from: The Register , Marcin Wichary .
