← Back to Home
AI NEWS 2026-11-22

AI news, 22 November: AI agents find and exploit medical software flaw in minutes

AI agents discover and exploit zero-day in 11 minutes. Amazon Threat Intelligence reported that autonomous AI agents, directed by a China-linked actor using the open-source Cairn framework, independently found and exploited a previously unknown critical vulnerability in MONAI Label, an open-source tool for AI-assisted medical image annotation. Given only a high-level goal to obtain shell access, the agents mapped the application and achieved arbitrary file writes in about 11 minutes, later gaining root control of a university server in East Asia. Amazon assessed it as a zero-day with high confidence; a fix has been merged but no full patched release was immediately available. The disclosure, covered by VentureBeat and others, highlights how AI can accelerate cyber attacks far beyond traditional speeds.

UN talks on ‘killer robots’ end without binding rules. Week-long United Nations discussions in Geneva under the Convention on Certain Conventional Weapons concluded with agreement to elaborate an international instrument on lethal autonomous weapons systems by 2028, according to Reuters and The Guardian. Nearly 100 countries sought a legally binding treaty soon, but opposition from major powers including the US, Russia and Israel produced a weaker consensus-based process. Campaigners called the outcome disappointing, while the Swiss chair described it as progress after more than a decade of talks amid growing use of such systems in conflicts.

Cursor explores phasing out Chinese open-weight models. SpaceX-owned Cursor told two US House committees it plans to explore dropping People’s Republic of China-developed open-weight AI models over geopolitical risk, a House Homeland Security and China Select Committees report stated, as reported by CNBC. Cursor’s own Composer coding models were built on Moonshot AI’s Kimi base. The report also examined use by companies including Airbnb and DoorDash and recommended broader restrictions on Chinese models in federal systems.

Trump threatens tax on overseas AI data centres. US President Donald Trump said the United States would not allow companies to build data centres abroad and would impose a “tremendous tax” on those that do, aiming to keep AI infrastructure domestic and counter China, according to Reuters. He offered few details on implementation and noted the idea was new even to some of his team. The remarks come amid domestic pushback against data-centre expansion over energy and local impacts.

Signal blocks AI assistants from reading chats. Encrypted messaging app Signal restricted screenshots on iOS and Android by default to stop operating-system AI agents such as Siri and Gemini from capturing private conversations, calling the trend an “existential threat” to its privacy guarantees, according to Reuters and The Verge. Users can opt out, but Signal warned this would undermine protections. The company urged Apple, Google and Microsoft to give developers better controls.

Thinking Machines releases compact open model for laptops. Thinking Machines Lab launched Inkling 2 Mini, an open-weight mixture-of-experts model with 26.4 billion total parameters that loops middle layers for stronger reasoning while fitting in roughly 20 GB, enough to run on many laptops. It leads comparable open models on several agentic coding benchmarks, the company said.