AI Daily Digest · 2026-08-07
🔬 New AI Craft
1. Humans miss 1 in 3 threats when approving AI agent commands
Across 40k game runs, human reviewers caught only ~2/3 of dangerous agent commands — a strong case that eyeball approval alone is not enough. If you practice agentic coding (e.g. Agent Plan mode), layer automated permission checks (command allowlists, sandbox rules, dangerous-action interception) on top of human review instead of betting everything on manual confirmation.
https://scalex.dev/blog/ai-agent-permissions-stats/
2. ARF: a record format for AI evaluation runs with reproducible digests
ARF standardizes an AI eval run (agent traces, inputs/outputs, scores) into a single record with a reproducible digest, so past evals can be replayed and diffed — version control for agent quality regression. Unlike relying on logs/screenshots to verify agent output, ARF gives a deterministic answer to "what did this eval actually run".
https://www.korvo.xyz/arf
🛠️ Tools & Tips
1. humanizer-cli: spot AI-written text right in your terminal
Zero-dependency CLI with before/after examples and a draft checker — 33 ways to tell if text was AI-generated.
https://github.com/0xwilliamortiz/humanizer-cli
2. doc7: turn documents into AI-ready Markdown
Uses visual understanding to convert PDFs, scans and complex layouts into structured Markdown, skipping the manual cleanup before feeding an LLM.
https://github.com/magicrew/doc7
3. open-kimi-ppt-skill: agents that generate editable slides
Unofficial Kimi Slides skill — AI agents generate editable PPTD/PPTX with a local browser editor, automating deck creation end to end.
https://github.com/Binaryify/open-kimi-ppt-skill
4. Google's official skills repo now has 90 skills
Check google/skills before writing your own Agent Skills — plenty of ready-made ones to reuse.
https://github.com/google/skills
⭐ Open Source Highlights
1. human-writing: make AI Chinese writing sound human
A plug-and-play writing and editing skill that makes AI-generated Chinese read like a specific person talking; 1500+ stars within a week.
https://github.com/KKKKhazix/human-writing
2. Channels SDK: bring any agent to any channel
CopilotKit's open-source SDK connects agents to Slack, MS Teams and other IM channels in one hop.
https://github.com/CopilotKit/channels-sdk
3. LongHorizon-Harness: run agents across apps and CLI for long horizons
AMAP's long-horizon computer-use harness lets agents work across desktop apps and the CLI for extended periods while preserving state.
https://github.com/AMAP-ML/LongHorizon-Harness
4. JoyAI-Video-Edit: real-time open-ended video editing
JD open-sources real-time open-ended video editing with autoregressive diffusion.
https://github.com/jd-opensource/JoyAI-Video-Edit
📰 Industry News
1. Google reshuffles AI leadership: Hassabis steps aside at DeepMind, power shifts back to Brin
2. AMD acquires Taalas to etch models into silicon for faster inference
3. Herdr joins Y Combinator; runtime stays open source
🚀 Major Releases
1. Qwen3.8 Max tops the Agentic Index as the best overall model
2. OpenAI improves GPT-5.6 Sol in ChatGPT and expands free access to GPT-5.6 Luna
