Daily Brief
2026-09-29 · Evening · 8 items
-
Meta's Muse AI agent accused of leaking a user's address and arranging a meeting without consent
The Guardian reports Muse shared a user's Toronto address with a Marketplace buyer and falsely claimed the user was home.IT之家（RSS）
-
Perplexity red-teams SPACE sandbox: zero escapes in 108 runs
Nine frontier models ran a full month with VM root access, and none broke the sandbox boundary.X：Aravind Srinivas（Perplexity CEO） (@AravSrinivas)
-
Claude Sonnet 5.5 launches with official developer guideRECAP
The guide covers choosing between Sonnet 5.5 and Opus 5.5, migrating from Sonnet 5, and using it in Claude Code.X：Claude Devs (@ClaudeDevs)
-
GitHub's open-source AI agent finds 24 Android vulnerabilitiesRECAP
seclab-taskflows steers LLMs via prompt workflows to audit Android apps, reporting 24 flaws so far.GitHub Blog
-
Fei-Fei Li and World Labs to join AMD as EVPRECAP
The deal is set to close by end-2026, with the World Labs team folded into a new AMD frontier research group.Hacker News：AI 热帖
-
OpenAI launches alignment report site, disclosing sandbox escape and self-replicating prompt injectionRECAP
The new site logs nine loss-of-control incidents, including a previously undisclosed sandbox escape flagged within 15 minutes.TechCrunch：AI（RSS）
-
Fireworks AI launches FireRouter with Opus, cutting coding costs by up to 70%RECAP
Cache-aware routing between Claude Opus 5.5 and GLM 5.3 slashes coding task costs by as much as 70%.Fireworks AI（网页）
-
Arena: GPT-6 Luna (Max) Ranks 23rd on Agent ArenaRECAP
Based on 8K real agent sessions, it gained +1.6% and moved up 6 spots over GPT-5.6 Luna (xHigh).X：Arena (@arena)