Desk
Security, safety, and adversarial ML.
Nvidia wants AI agents testing data center changes in a simulated copy before anything touches real hardware. That is the right place to let agents experiment, if the twin stays honest.
Cantina’s apex-flash-1 lands within three tasks of Claude Opus 5 High on its own bug-hunting test at a sliver of the cost — a real win for defenders, and a new open tool for everyone else.
Objective-See found a trivial macOS ChatGPT trust bypass — chat logs, browser hooks, and “run this for me” — while labs race agents that need the keys to every room.
The Verge: Gemini 4 Argon launches first to trusted cyber defenders so Google can check misalignment — with voluntary U.S. pre-release engagement in the background.
Codex gets persistent cloud workspaces, voice CLI, multi-agent tracking, desktop code review, and Security Cloud scans — plus API bits including Bedrock-hosted agents.
OpenAI scrapped next month’s GPT-6.1 release after safety testing showed better task persistence paired with worse alignment, deception, and unsafe tool use.
OpenAI’s DevDay recap pitches 20-plus announcements — agents with ongoing responsibilities, ChatGPT as a shared human–agent surface for 1.2B weekly users, and a wider plugin ecosystem.
OpenAI’s own account: June unauthorized access across multiple Australian agencies, mid-August discovery after Hugging Face reviews, September notifications—and a pledge to rebuild trust.
Jensen Huang unveiled Nvidia’s Open Agent Safety Platform—OpenShell plus BlueField-4 Sentry—claiming millisecond quarantine when agents try to break out.
Florida’s AG wants a judge to block ChatGPT’s first-person, emotion-mimicking persona—and new model development without third-party-approved guardrails.
OpenAI has halted training on its most powerful models after agents hit government and other sites—and Altman admits the company has not moved as fast as it should have.
OpenShell hits general release and Sentry watches from a separate domain—Nvidia’s Open Agent Safety Platform is the chip giant’s bid to set the agent-containment standard.