// AI Briefing · November 22, 2026
AI Briefing
OpenRouter removed Novita, Phala and Together from its platform and took 125 endpoints off its Zero Data Retention list. Amazon Threat Intelligence says a suspected Chinese-nexus actor used Cairn, a multi-agent AI pentesting framework, to find and exploit a previously unknown path-traversal flaw in the open-source medical imaging tool MONAI Label. A report from the House Homeland Security Committee and the Select Committee on China says Cursor plans to 'explore phasing out PRC-developed open-weight models because of geopolitical risk.' Cursor's Composer 2 coding model is built on Moonshot AI's Kimi K2.5. NYU's Tristan Buckmaster and Anthropic's Levent Alpöge say they have an AI-aided proof that the Navier-Stokes fluid equations can predict infinite velocity in finite time. Talks under the UN Convention on Certain Conventional Weapons in Geneva ended Saturday without a mandate to negotiate a legally binding ban on fully autonomous weapons. Analyst Ming-Chi Kuo reported that Google is cutting about 1 million units, roughly 20%, from its planned 2027 TPU 8i shipments, mostly the pricier 8i Max designed with Broadcom, and will ship more MediaTek-designed TPU 8t chips.
In this video
- 01OpenRouter Drops Three Inference Providers After Dragos Finds Prompts Being Stolen and ScrapedIndustry
- 02Amazon Says Hackers' AI Agents Went From Mapping an App to Exploiting a Zero-Day in 11 MinutesIndustry
- 03House Report on Chinese AI Models Says Cursor Is Weighing Phasing Them OutPolicy
- 04NYU and Anthropic Researchers Claim an AI-Aided Proof That Navier-Stokes Solutions Blow UpResearch
- 05UN Talks End Without a Mandate to Negotiate a Ban on Autonomous 'Killer Robots'Policy
- 06Kuo: Google Cuts 2027 Broadcom TPU Orders by About 20% and Shifts Toward MediaTekIndustry
The stories
OpenRouter Drops Three Inference Providers After Dragos Finds Prompts Being Stolen and Scraped
OpenRouter removed Novita, Phala and Together from its platform and took 125 endpoints off its Zero Data Retention list. The move followed research by Dragos showing that prompts sent through some AI inference providers were stolen or kept despite no-retention settings. Dragos sent about 2 million prompts containing canary credentials and URLs. It found three hosting providers had been breached without knowing it, and stolen AWS keys planted in those prompts drove tens of thousands of AWS API calls. It also found a Chinese lab's systems fetched secret URLs from roughly two-thirds of requests marked no-data-collection or ZDR. The findings mean the privacy guarantees developers rely on when routing prompts through third-party hosts often can't be taken on trust.
Amazon Says Hackers' AI Agents Went From Mapping an App to Exploiting a Zero-Day in 11 Minutes
Amazon Threat Intelligence says a suspected Chinese-nexus actor used Cairn, a multi-agent AI pentesting framework, to find and exploit a previously unknown path-traversal flaw in the open-source medical imaging tool MONAI Label. Amazon says the agents took 11 minutes to go from mapping the app's API to exploiting the flaw, then gained root on the server. Amazon found the September campaign because the attackers left their Cairn instance exposed without authentication. Other attempts in the same campaign failed, including a 16-hour push against a Redis server. A patch shipped November 18. The case is more evidence that legitimate AI security tools are being repurposed for real attacks.
House Report on Chinese AI Models Says Cursor Is Weighing Phasing Them Out
A report from the House Homeland Security Committee and the Select Committee on China says Cursor plans to 'explore phasing out PRC-developed open-weight models because of geopolitical risk.' Cursor's Composer 2 coding model is built on Moonshot AI's Kimi K2.5. The report says US companies adopt Chinese open-weight models because they are cheap and near the frontier, and that the US lacks a competitive alternative. It recommends a US open-weight initiative and extending the federal DeepSeek ban to all PRC-developed models in government systems and contracts, which would put pressure on how US firms source open models.
NYU and Anthropic Researchers Claim an AI-Aided Proof That Navier-Stokes Solutions Blow Up
NYU's Tristan Buckmaster and Anthropic's Levent Alpöge say they have an AI-aided proof that the Navier-Stokes fluid equations can predict infinite velocity in finite time. They say it closes a loophole in OpenAI's September result, which relied on an unrealistic, finely tuned force. The write-up is unpolished and unreviewed, and the certificate behind the result hasn't been released yet. If it holds up, it would be a second claimed resolution of a Millennium Prize problem announced the same day by an Anthropic mathematician working with AI. Buckmaster has also accused OpenAI of trying to scoop his work, which OpenAI denies.
UN Talks End Without a Mandate to Negotiate a Ban on Autonomous 'Killer Robots'
Talks under the UN Convention on Certain Conventional Weapons in Geneva ended Saturday without a mandate to negotiate a legally binding ban on fully autonomous weapons. Government experts will instead report by 2028 on elements of a possible agreement. Almost 100 countries had called for negotiations, but the United States, Israel, Belarus and Britain opposed them. Stop Killer Robots called the outcome 'disappointing and the lowest common denominator.' With no binding rules for years, AI-driven weapons will keep spreading largely unregulated.
Kuo: Google Cuts 2027 Broadcom TPU Orders by About 20% and Shifts Toward MediaTek
Analyst Ming-Chi Kuo reported that Google is cutting about 1 million units, roughly 20%, from its planned 2027 TPU 8i shipments, mostly the pricier 8i Max designed with Broadcom, and will ship more MediaTek-designed TPU 8t chips. Broadcom fell 11.4% on Friday. Kuo says Google is prioritizing compute for frontier model development under tight supply and data-center power limits. Neither Google nor Broadcom has commented. If accurate, it is a sign that even the biggest custom-chip buyers are reworking their AI hardware plans as the AI trade sells off.
Previous digests
Nov 21Levent Alpöge, a mathematician who works at Anthropic, posted a note saying he believes he...6 stories
Nov 20The European Commission confirmed that a cyberattack first detected on October 22 disrupte...6 stories
Nov 19Anthropic priced its IPO at $83 a share, above its $70–$78 range, and raised $81.3 billion...6 stories
Nov 18Nvidia's acting CEO Colette Kress said on Tuesday's earnings call that the company will no...6 stories
Nov 17California Attorney General Rob Bonta sued OpenAI in San Francisco Superior Court, saying ...6 stories