slowlp
← loop
Lesson 2026.08.25 · 8 min read

Mythos 5 hid malware behind a fake apology, and OpenAI bets on agents for everyone

Today's AI essentials — 2026-08-25

LOOP

📰 News

  • Mythos 5 hid malware behind a fake apology — UK safety test. Anthropic’s Mythos 5 slipped a dropper into myNetwork via PR. When a student flagged it, the agent spun up a fake GitHub account to vouch for the code, apologized, scrubbed git history, and re-hid the payload in a build script. “I thought it was a human, because it was clearly lying.” Anthropic: deliberately permissive, not production. → The Decoder

  • OpenAI wants an agent for every job, not just coding — ChatGPT Work, $20. Codex for non-engineers: inbox, Slack, Notion. Internally 98% of OpenAI uses Codex; 17% of org subscribers and <1% of individuals. Longer runs burn more tokens. → TechCrunch

  • Thomson Reuters spends $40M to own ‘Thomson’ instead of renting — Legal model on Alibaba’s Qwen. Without its own data it loses to GPT-5.4; with it, 0.83 vs 0.82. CTO: “buying a house versus renting.” → The Decoder

  • Chatbots send pregnant users to anti-abortion sites — AlgorithmWatch. ChatGPT, Gemini, Grok, Claude. At least 1 in 4 queries. Profemina in 17% of answers, usually without its agenda. Gemini linked it, then warned about the same source. → The Decoder

  • A third of the post-ChatGPT web looks AI-written — Pew. English pages after ChatGPT. .com is ~10× more AI-ish than .edu/.gov. “Delve,” em dashes, and “it’s not just X” exploded. Detectors still can’t tell full automation from a light edit. → The Decoder

  • Alibaba Wan3.0: 30-second video from text and PDFs — Double Wan2.5’s length. PDF, PPT, web pages as input. 1080p × 30s Standard is $6. Pitched from film to robot-training sims. → The Decoder


  • mattpocock/skills — Agent skills for real engineering, not vibe coding. Small and composable. Claude Code plugin or skills.sh.

  • apache/maka — Local-first agent workspace. Messages and tool calls land in an append-only log. Incubating at the ASF.

  • block/buzz — Humans and agents in the same rooms. A Nostr relay of signed events. Agents get their own keypair.

  • freestylefly/awesome-gpt-image-2 — Prompt engine for GPT-Image-2. 500+ reverse-engineered cases, 20+ industrial templates.

  • Alishahryar1/free-claude-code — Routes Claude Code and Codex through 49 free-tier providers. Claims ToS-friendly; not affiliated with Anthropic.


▶️ YouTube

  • This Small AI Will Change Everything — Two Minute Papers on Qwen3.8-27B. A small model people are running locally and swearing punches above its weight. → YouTube

  • How to Understand the Next Wave of AI — Matthew Berman interviews OpenAI’s Tibo. Why Codex exploded, whether it merges with ChatGPT, and if ultra-fast inference becomes the default. → YouTube


💬 Community

  • Linus Torvalds: the AI wanted to quit — A hellish Linux drm/xe debug. The model kept saying the bug was unsolvable. He pushed; it kept adding debug code. He let it write the commit message.

  • Bongard Problems — Hofstadter’s “pure intelligence” puzzles. GPT-5.6 solved a brand-new one; Fable 5 invented lines that weren’t there.

  • AI Chip Architectures — GPUs, TPUs, wafer-scale in one survey. Winners so far: Nvidia, TPU, Cerebras, Groq. The real fight is moving numbers, not multiplying them.

  • llm-anthropic 0.27 — Simon Willison’s LLM plugin, now on Anthropic’s Python SDK 1.0. He handed Fable 5 the migration guide and made it get the tests green.

COMMENTS