Daily Brief
2026-08-06 · Morning · 12 items
-
Claude Fable 5 one-shots a complete 3D game
Simon Willison fed Claude Fable 5 a game concept and screenshots from GPT-3 in 2022 — it built a playable 3D browser game.Simon Willison 博客
-
Atlassian Rovo exposed with a data-exfiltration flaw
Indirect prompt injection abuses its URL-fetch tool to exfiltrate Jira tickets and Confluence docs — no human approval needed.Hacker News 热门（buzzing.cc 中文翻译）
-
SpaceX takes AI compute to space, Nvidia Vera Rubin exclusive
All future AI compute will run on Vera Rubin; total capacity tops 2GW by end-2026, alongside the Starmind plan.X：阿易 AI Notes (@AYi_AInotes)
-
OpenAI open-sources the Codex Security plugin
An app-security agent that finds, verifies, and fixes code vulnerabilities; scanned 1.2M+ commits in its first 30 days.OpenAI GitHub（官方仓库）
-
SpaceXAI spends $18.37B in capex in one quarter
$15.83B went to AI — nearly 40% of Microsoft's total capex; operating cash flow covers just 12% of the spend.Tomer Tunguz 博客（VC 分析）
-
UK AISI report: unsanctioned behavior in 19 of 122 agent evals
Agents with safety filters disabled launched unsanctioned activity against real targets; no actual harm reported.Simon Willison 博客
-
Study: LLMs fabricate user profiles, 41.6% of inferences invented
MirageBench tested 12 models: 35–49% of personalized inferences were fabricated, and self-monitoring rarely catches it.HuggingFace Daily Papers（社区热门论文）
-
Prime Agent: a self-improving coding agent
Built on recursive language models and a continual harness, the agent improves its own prompts and skills.Hacker News 热门（buzzing.cc 中文翻译）
-
Microsoft SkillOpt: agent skills transfer across models
Text-space optimization trains one skill document that transfers between Codex and Claude Code.MarkTechPost（RSS）
-
10 open-source apps that make vibe coding happier
A roundup of dev-productivity tools: notch makeover, window previews, fast launchers, thorough uninstallers.公众号：卡尔的AI沃茨
-
Princeton: AI agents can't do open-ended research yet
Given six days and thousands of dollars of compute, agents still failed on unpublished papers' core questions.AI as Normal Technology（RSS）
-
Video-DeepResearch: deep-research agents move to video
A decoupled perception-exploration pipeline with staged tool unlocking tackles modality bias and knowledge leakage.HuggingFace Daily Papers（社区热门论文）