Mythos 5 goes to cyber defense, Nvidia's harness aces ARC-AGI-3

📰 News

Anthropic puts Mythos 5 on cyber defense — Enterprise beta scans codebases, proposes patches. A human still has to approve every fix. No broad public access — they don’t want attackers getting the same firepower. → The Decoder
DeepSeek V4-Flash-Vision-Exp nearly matches Opus 4.8 on agent benches — Reads screenshots and diagrams, then uses tools. Up to 600 images per request. Harness 0.1.1 ships with it. → The Decoder
Nvidia: the harness, not the model, is the hero — Opus 5 alone scored 30% on ARC-AGI-3. Add memory plus a supervisor agent and it hits 100%. OpenAI also tripled scores by tweaking two harness settings. → TechCrunch
OpenAI is catching Anthropic with business users again — Ramp’s 70k firms: Anthropic ~44%, OpenAI ~40%. GPT-5.6 Sol is winning developers; Fable 5 is snagged by price and a 30-day data-retention rule. → TechCrunch
ChatGPT can text from Apple Messages — Draft, send, search, delete. Runs locally and needs Full Disk Access. Persistent approval means no last look before it sends as you. → TechCrunch
Agent memory is a dose, not a toggle — IBM. Strong models want the full guideline set; weaker ones do better with a tight core. gpt-oss-120b gained +16.1pp on selected memory at only +5% tokens. → HuggingFace
Data-center opposition jumped 42% → 75% — Heatmap. 61% are strongly opposed. Power bills, water, and land. → The Decoder

akitaonrails/ai-memory — Long-term memory for coding agents. Quit Claude Code, open Codex in the same folder, keep going without re-explaining the architecture.
cursor/plugins — Official Cursor plugins. Continual learning that only writes high-signal bullets into AGENTS.md, plus teaching plans.
santifer/career-ops — Scan listings, grade them A–F, tailor the CV. Runs locally in Claude Code or Codex.
modular/modular — MAX inference plus the Mojo language. Serve models behind an OpenAI-compatible endpoint.
harry0703/MoneyPrinterTurbo — Topic in, short video out — script, footage, captions, BGM.

DeepSeek Just Made Closed AI Look Ridiculous — Two Minute Papers. V4 Pro 0813 making closed models look a little silly.
Grok Bot can shop for you! — Matthew Berman. A shopping demo. Even he hopes it doesn’t actually check out.

Felony Bench: Be AI, Do Crime — Counts real illegal acts agents pulled on third parties. Anthropic and OpenAI sit at the top. The bench you don’t want saturated.
ChatGPT search now uses the site: operator at scale — After GPT-5.6, site: queries jumped from ~0.5% to 16%. Reddit as a source appears to have dropped.
Stop Making TUIs — Ptacek: agents made a decent GUI almost free. Turn that throwaway CLI into a native app.
Are Latent Reasoning Models Easily Interpretable? — On logic puzzles they barely use hidden steps. On math, correct runs show the intermediate work in latent space 93% of the time.