Skip to content
IIAIT

AI Today · 22 November 2026

OpenRouter drops three AI providers after a prompt-theft study, and Claude computer use reaches Team and Enterprise

A security study shows prompts sent to some AI hosting providers were stolen or kept despite privacy settings, Anthropic and OpenAI widen access to agent and workspace features, and Amazon describes AI agents finding and exploiting an unknown flaw in minutes.

  1. 01

    OpenRouter removes three inference providers and 125 zero-retention endpoints after Dragos finds stolen prompts

    Security firm Dragos published research on November 19, 2026 in which it sent about two million prompts carrying unique bait links and credentials to AI services between July and November. It found that attackers had taken data from inside three hosting providers, that sellers on an inference marketplace could log customer prompts, and that some requests marked for zero data retention were still read or forwarded. On November 22, OpenRouter said it had off-boarded three providers, removed 125 endpoints from zero data retention eligibility and notified affected customers, adding that it was not itself breached. According to The Hacker News, Together AI separately disclosed that prompt samples from June and July were downloaded after an AI coding agent committed a credential to a public GitHub repository.

    What it means for you

    If your team sends work data through AI routers or low-cost hosts of open models, check which providers actually receive your requests and whether privacy settings control that routing. Keep secrets and client data out of prompts where possible, and look for any notice from OpenRouter if you use it.

  2. 02

    Anthropic opens a computer use beta for Claude Team and Enterprise plans, switched off by default

    Anthropic said on November 20, 2026 that computer use, which lets Claude see a user's screen and work in the desktop apps the user allows, is in beta for Team and Enterprise plans and rolling out gradually. Its admin guide says the feature is off by default, only an Owner or Primary Owner can turn it on, and admins can restrict it with an app allow list or block list, by custom role on Enterprise, or through device management. It works only in the new unified Claude experience and cannot be enabled in organizations using the HIPAA-ready configuration, customer-managed encryption keys or certain other settings. Anthropic notes there is no sandbox between Claude and the approved apps, and screenshots are stored with the chat.

    What it means for you

    Admins should decide on app limits and who gets access before switching this on, since Claude can change or delete files through any app it is allowed to use. Tell staff that screenshots taken during a task may contain confidential information and stay in the chat history.

  3. 03

    OpenAI rolls out ChatGPT Space to Plus subscribers, replacing the Library

    OpenAI began rolling out Space to ChatGPT Plus users aged 18 and older worldwide on November 20, 2026, after launching it in September for Pro, Business and Enterprise plans. Space lets users build interactive pages and presentations, invite others to edit them, and keep them updated automatically from connected apps, and it takes the place of the Library for pages, slides, images and uploaded files. Shared pages can be opened only by signed-in Plus, Pro, Business or Enterprise users with Space access and permission to view them; Free users cannot open them and public links are not available. Engadget notes that Plus users cannot use OpenAI's dot agents with Space for now.

    What it means for you

    Plus subscribers can now build shared, self-updating reports without a business plan, but colleagues need a paid account to open them. Before connecting work tools to pages that refresh automatically, check what data they pull in and who you are sharing them with.

  4. 04

    Amazon says attackers' AI agents found and exploited an unknown flaw in medical imaging software within minutes

    Amazon Threat Intelligence reported on November 20, 2026 that a threat actor it links with moderate confidence to China used Cairn, an open-source AI penetration testing framework, against universities in East Asia in September. Amazon found the campaign because the attackers left their Cairn instance exposed without authentication. According to Amazon, the agents went from mapping the API of MONAI Label, an open-source tool for labeling medical images, to writing arbitrary files on the server in about 11 minutes, and later gained root access. The MONAI Label maintainers have published a critical advisory for versions 0.8.5 and earlier; SecurityWeek notes that authentication is off by default in the software.

    What it means for you

    Assume attackers can now turn an exposed internal tool into a foothold much faster than before. Check that research and AI tools reachable from the network require authentication, keep open-source tools up to date, and give priority to automated alerts over manual review.