AI digest — 2026-09-27
Top line: A genuinely quiet window — six items, nothing new to this digest. The only adds are quantifications and comparison posts on stories already on the record; yesterday's training-pause fallout remains the beat's live story.
One-liners
- Security roundups pair: "LLM-jacking" — hackers hijacking AI accounts and servers to steal expensive compute for automated cybercrime (the420.in) — the "steal the key, not the money" economics of the agentic-attacks thread, now with its own name; and a crypto-crime report counting $2.2B in 2026 hacks across 288 attacks, with AI-assisted attacks as a named category (yellow.com). Corroboration layer; no new mechanism.
- Model-comparison posts: note.com pair on GPT-6 Luna's limits ("common tests") and Opus 5.5 vs GPT-6 Sol (the same-day releases compared on price-per-intelligence) — hobbyist/practitioner color on the Sep 22 launches; nothing benchmark-grade.
- Sarvam AI's Saaras V4 — full writeup (marktechpost) of the Sep 25 search-item: STT across all 22 Indian languages + global English, keyterm prompting, sub-150ms streaming. Regional-model color;
efficient-tieradjacent.
Dropped
- XDA "I stopped my local LLM agent from running out of context" — hobbyist how-to; the local-agent genre, not news.
- The window's genuinely new information sits on the X side: a flagged claim that OpenAI and Anthropic are investigating "tens of thousands of incidents — not dozens… could grow well beyond" (meme-account provenance via Musk RT; see the ai-social digest) — held out of the storyline until anything primary or press-grade corroborates the scale.