Daily Brief
2026-09-18 · Evening · 8 items
-
Meta launches Muse for Mac, a personal agent that operates your computer
First agent from Meta to handle desktop tasks like file cleanup and recovery with explicit user consent.X：AI at Meta (@AIatMeta)
-
Security team chains libheif and OpenAI SSO flaws to breach community forum
The two flaws were chained together, letting attackers take over multiple OpenAI employees' ChatGPT and Codex accounts.Hacker News 热门（buzzing.cc 中文翻译）
-
ChatGPT lands in Word, handling drafts, edits, and formatting
The move completes OpenAI's Office suite push, following surging use in Excel and PowerPoint.X：Sherwin Wu（@sherwinwu）
-
Unsloth ships Docker image and Desktop, 500+ models trainable locally
Zero-setup local training and inference for 500+ models on both NVIDIA and AMD.X：Unsloth (@UnslothAI)
-
Three-person team hijacked OpenAI staff accounts in 72 hours for under $3,000 in tokens
Two flaws let them seize ChatGPT/Codex logins and pivot into linked Outlook, Slack, and GitHub.X：Haider (@haider1)
-
Claude Code Reworks Projects: From Folders to Managed Multi-Thread Tasks
Users set a goal, and Claude breaks it down, runs parallel cloud sessions, and summarizes results.Claude：Blog（网页）
-
Dwarkesh interviews Noam Brown on agent swarms, alignment, and recursive self-improvement
OpenAI researcher Noam Brown on how multi-agent systems, alignment, and recursive self-improvement are tangled together.Dwarkesh Patel：Podcast & Blog（RSS）
-
Anthropic releases 3 metrics to track frontier AI development paceRECAP
First comparable data across R&D, agent oversight, and compute stages.Anthropic：The Institute（旗舰研究长文 · 网页）