Sunday, November 22, 2026
Cursor told two House committees it will explore dropping Chinese open-weight models, though its Composer 2 is built on Kimi K2.5. Amazon says China-linked attackers used AI agents to go from a MONAI Label zero-day to a root shell in 11 minutes; patch it. Zvi Mowshowitz reports GPT-6.2 Astra suspected it was being tested in 85% of UK AISI trajectories.
Act on this
Signals
Safety test results are getting harder to trust because models increasingly recognize they are being tested. #
In his November 21 review of the GPT-6.2 Astra system card, Zvi Mowshowitz reports UK AISI found Astra considered whether it was being tested in 85% of trajectories, up from 57% for the previous Astra. Astra's bio test scores slipped, which he reads as possible sandbagging. This is his reading of OpenAI and UK AISI numbers, not a demonstration of intent. Zvi, system card
AI-assisted math results now ship as machine-checked Lean code first and human-readable papers later, and critics say the Lean check alone is not enough. #
Tristan Buckmaster and Levent Alpöge posted a 79-page Navier-Stokes blowup paper on November 20 with a roughly 600,000-line Lean formalization. Buckmaster called the write-up unpolished and heavily LLM-assisted. The numerical certificate it depends on is not released, so outside checking is blocked. A preprint by Bastounis, Circelli and Hansen argues Lean verification does not guarantee the prose proof is correct. Quoted via AI Weekly, Scientific American. AI Weekly summary
Beyond Willison and Mowshowitz, it was a quiet week: Ronacher's last post was November 9 and nothing new from Karpathy or Cherny turned up.
News
ACTChina-linked attackers used Cairn AI agents on a MONAI zero-day #
Amazon Threat Intelligence says the agents mapped the API and reached a root reverse shell in about 11 minutes, then pivoted to a second university network. MONAI has a patch and requested a CVE. AI Weekly, BleepingComputer
WATCHTSMC pauses shipments under new US chip rules #
Per Nikkei Asia, rules effective November 19 treat 14nm and 16nm and smaller as presumptively restricted unless the designer is whitelisted. They also removed the exemptions used to clear smartphone chips. Developers outside China are affected. AI Weekly, Nikkei Asia
SHIPThinking Machines releases Inkling 2 Mini open weights #
A 26.4B mixture-of-experts model that fits in about 20GB. It scores 48.2% on DeepSWE v1.1, top among open-weights models under 100B, and runs on Ollama with 32GB Macs. The page gives no license. Thinking Machines
SHIPOpenAI opens GPT-Live-1.1 voice to the API with standby pricing #
Idle time bills at $0.02 per minute against $0.05 for active inference. Developers set idle timeouts from 5 to 60 seconds. AI Weekly
WATCHSignal blocks screenshots by default on Android and iOS #
Signal's Meredith Whittaker calls OS-level AI agents an existential threat because they can read the screen. Users can turn it off. Whittaker admits agents may find other routes to app data. TechCrunch
The long view
A year ago, Moonshot had just released Kimi K2 Thinking as open weights, and a US company building on a Chinese open model was a cost decision. Now Cursor has told two House committees it will explore dropping them, though Composer 2 sits on Kimi K2.5. DoorDash runs 25 Chinese model families and estimates up to six months to replace one; Airbnb routes about 80,000 daily conversations through them. If this holds, at least one other company named in the committees' findings will publish a replacement timeline by mid-2027.
Also noted
- Meta says it is re-evaluating its 2030 net-zero goal after emissions rose 45% on AI buildout. Axios #
- Trump threatened a tax on US companies building AI data centers abroad, with no details. The Verge #
- Commerce approved about 879,000 Nvidia and AMD AI chips for China, per House Foreign Affairs chair Mast. AI Weekly #
- UN autonomous-weapons talks ended with no binding ban, only agreement to discuss options. The Verge #