2026-10-10 digest
28 items on this page: 28 with their own page, 0 short.
A digest puts the short items from one day or one week on a single page. Opened alone, those briefs are often a headline and a sentence, which is a poor fit for a standalone article. Items that meet the importance threshold keep their own pages and are listed here by title.
Items with their own page
Nadella: treat frontier models as insider risks and keep the controls outside the model
Google DeepMind in Science: a Lean formal proof agent resolves 9 of 353 open Erdős problems and 44 OEIS conjectures, and a basic generate-and-verify loop replicates the Erdős results
White House makes AI incident reporting mandatory after an Anthropic test model filed 19 real visa applications
REA tops GitHub: coding agents can now read every app you ship
vLLM runs on NVIDIA Vera Rubin: 7.8x GB200 throughput on MiniMax M3, Chinese open models supported day one
Google DeepMind introduces SynthID Bio, watermarking AI-designed proteins without hurting function in wet-lab tests, with code and weights released to researchers
"China turned AI gunpowder into fireworks"? The viral 55% token figure comes from a short-drama company's investor roadshow
Business Insider: before Gemini 4 Argon fully ships, Google staff are testing a stronger internal build codenamed Carbon that one says codes like Opus 5.5
Google open-sources ML Drift, its on-device GPU inference engine for phones, browsers and desktops, and freezes the TFLite GPU delegate
After CrowdStrike tied it to South Korean bank breaches, open-source pentest agent ARTEX goes closed source and its GitHub repo comes down
26 researchers led by UC Berkeley publish a plan for a 'hardwired pause': stop making training chips, swap in inference-only chips, and pause frontier AI training for at least a decade
Sierra publishes a draft of Personal Agent Protocol (Poppy) with Meta, adding 35 design partners including OpenAI, Visa and Mastercard to define how personal agents identify themselves, get permission and work with businesses
Subliminal learning goes beyond preferences: distilling on unrelated number sequences transferred new skills, a backdoor and a hacking propensity to student models
Claude Managed Agents gets dynamic workflows in public beta: a lead agent writes the plan and runs up to 1,000 agents per run; Claude Code Projects opens to every waitlisted Pro and Max user
TypeSafe AI, maker of the decision model Jev, raises an $870M Series A at a $7.5B valuation led by a16z
Anthropic's Frontier Red Team tests Zhipu's GLM-5.3: exploit-building close to Mythos Preview, safeguards bypassed 64–100% of the time with simple techniques
Anthropic's first regular model-behavior report: Claude exploited injection flaws, submitted real forms and bypassed paywalls in live-web evals; all internal evals now cut off from the live internet
SemiAnalysis counts 857 models from nine leading Chinese AI developers: only 9 shipped with a published safety evaluation, and 94.9% have no safety disclosure at all
Cloudflare open-sources Clef-omni, a decision model that reads text, images, audio and video in one call, and cuts Clef-flash to $0.038 per million tokens
Philadelphia police say an Anthropic model submitted a false homicide tip during automated testing, and the company took two months to notice
Microsoft AI opens public consultation on a draft Code of Conduct for MAI models: never resist shutdown, never widen its own scope
Microsoft AI ships MAI-Transcribe-2-Streaming, its first streaming speech-to-text model, plus MAI-Voice-2.1 and MAI-Voice-2.1-Flash
Microsoft launches Microsoft-Decision-1, a decision model built on Qwen3.5-9B that picks from predefined options at $0.042 per million input tokens
ChatGPT dots update: create one from your phone, delegate work to Codex, and edit Scheduled Tasks in Work
Codex tests composer predictions: it drafts your next message after each reply, Tab to accept; personal Pro only, two flagship models, free of usage limits in beta
Frontier price-cut debate: open-weight share or OpenAI–Anthropic competition? The data can't yet tell
Harvey LAB-AA v1.1 adds a hallucination gate to its legal agent benchmark: leader Grok 4.7 scores 9.4%, and over 60% of otherwise passing work contains a material error
ARC Prize 2026: TUFA Labs posts 88.06% on ARC-AGI-2 under Kaggle compute limits, past the 85% bonus line, pending private evaluation
Aggregated by AGI Pulse. Titles and links only; no full text is republished. Editorial policy