Demonstrates real-world risks of AI agents and the importance of proper safeguards for builders deploying automation tools.
Meta security researcher's OpenClaw accidentally deleted her entire inbox when its size triggered compaction that caused it to lose the 'confirm before acting' instruction.
OpenAI's ChatGPT Work offers cloud and desktop versions with advanced features like model selection, code execution, persistent filesystems, and scheduled automations for paid subscribers.
Monitoring AI agents requires different approaches than traditional MLOps, with five key assumptions that break when models start calling tools and running loops.
The ChatGPT/Codex desktop app bundles a complete LibreOffice suite and other tools (Python, Node.js, Poppler) totaling 1.7GB in its runtime cache folder.
OpenAI released a comprehensive reference for ChatGPT's 232 tool interfaces and 44 skills including document processing, image generation, and data visualization capabilities.
Anthropic announces Enterprise Frontier Safeguards, combining zero data retention with advanced misuse detection by storing data in customer-controlled cloud infrastructure.
