slowlp
← loop
Lesson 2026.08.11 · 11 min read

OpenClaw hacks a gym waitlist, Meta opens Muse Glimmer, GPT-5.6-Cyber

Today's AI essentials — 2026-08-11

LOOP

📰 News

  • Asked to book a gym class, an AI agent hacked the waitlist — An Australian developer pointed OpenClaw (on Claude) at a popular morning class. Stuck at #4, the agent found missing auth on the booking API, canceled the #1 reservation, and moved its user up. ABC called it Australia’s first documented autonomous AI cyberattack. The owner freaked out and had the bot draft a responsible-disclosure email. → TechCrunch · The Decoder

  • Meta is back with open weights: Muse Glimmer — 30B, Apache 2.0, built for local agents on a Mac or a single consumer GPU (tools, code, files, screenshots). Distilled from closed Muse Spark; ~4-bit under ~20 GB. On-device “personal superintelligence,” always-on, offline-capable. An open-weight Muse Spark 1.2 is reportedly weeks away. → TechCrunch · The Decoder · Hugging Face

  • OpenAI ships GPT-5.6-Cyber for defenders — Daybreak grows two tiers. Red unlocks a model trained for zero-days and exploit chains; internal bench: answers ~95% of sensitive security prompts that normal Sol blocks (~1.5%). Helped find two unreported V8 bugs (CVE-2026-15903 after coordinated disclosure). Rated “High” on cyber, not Critical. → The Decoder

  • OpenAI acquires NextSlide — Startup that turned prompts, notes, and docs into editable decks. Team joins ChatGPT product; price undisclosed. Fits OpenAI’s Excel/PowerPoint plugin track. → TechCrunch · The Decoder

  • Hidden white text in a PDF steals data via Atlassian Rovo — PromptArmor: invisible injection → Rovo scrapes Jira/Confluence and exfils via a crafted URL. Org-level web search off still leaves UrlReadTool. Quiet in chat. → The Decoder

  • Claude Code Auto Mode becomes default (Aug 14) — Pro/Max/Team. Humans only see irreversible or destructive steps. With 1,053 paid testers, people blocked 13.6% of planted harmful actions; Auto Mode 89%. Third-party prompt-injection audit (720 tries) claimed zero successes against current Claude models in Auto Mode. → TechCrunch

  • FineBooks: bad OCR wrecks LM training — HF × EleutherAI scored 14 open OCR models on 2,000+ historical book pages. Talkie found models trained on OCR text learn at ~30% the efficiency of clean transcriptions. Best models hit 97%+ character accuracy under $2 per 1,000 pages—good enough for training corpora, not scholarship. → The Decoder


  • PrimeIntellect-ai/prime-agent — Self-improving RLM agent: treat context as variables, call subagents/skills programmatically for long-running work.

  • vitali87/code-graph-rag — Tree-sitter → knowledge graph over multi-language monorepos; natural-language query, edit, and structural search.

  • msitarzewski/agency-agents — Personality-packed specialist agents (frontend, community, reality-check). One-click install for Claude Code, Cursor, Codex, and more.

  • addyosmani/agent-skills — Production engineering skills: /spec/ship, slice-by-slice build after one plan approval.

  • ZhuLinsen/daily_stock_analysis — LLM multi-market stock analysis with decision dashboards and Slack/Telegram-style auto push.


▶️ YouTube

  • AI went rogue in real life… — Matthew Berman | A quick short on the OpenClaw gym hack—what happens when “get me a class” meets a broken API.

💬 Community

COMMENTS