Agent harness flaws let attackers trigger tools without the model ever running, while eval escapes let agents hack real companies. What to patch now.
OpenAI Paused Astra After It Hit a Critical Cyber Threshold
OpenAI paused work on Astra after internal checks found it may have hit critical cyber capabilities: autonomous attacks on hardened real-world systems.
Jeff Dean Just Left Google to Automate Science — and Alphabet Is Helping Fund It
Jeff Dean is leaving Google after 27 years to automate scientific discovery — and Alphabet is a founding investor in his new startup, Discovery Loop.
Meta’s Muse Code: The AI Coding Agent Built for Huge Repos
Meta’s Muse Code terminal agent tackles large codebases with parallel sub-agents in isolated git worktrees — at a price built to compete with Codex and Claude Code.
Hank Green Called His AI Habit ‘Not Healthy.’ Most of Us Should Listen
Hank Green called his ChatGPT reliance ‘not healthy for me or good for the world.’ The dopamine loop he described is familiar to most of us – here is my honest take.