Tech news · 22 November 2026

Tech news, 22 November 2026: AI agents exploit a MONAI Label zero-day, an Entra flaw, a House report on Chinese open-weight models and Signal versus OS assistants

Amazon reported that a China-linked operator's autonomous AI agents found and exploited a zero-day in an open-source medical imaging tool. Microsoft disclosed that a flaw in Entra Connect Health let attackers read tenant metadata including administrator emails. Two House committees recommended limits on Chinese open-weight models, and Signal began blocking operating system AI assistants from reading chats.

Tech newsSumanta Mukhopadhyay6 min read6 sources
#news#security#ai#governance
TECH NEWS · TECH NEWS · 22 NOVEMBER 202622 Nov 2026 · 08:55 GST

Today's brief covers stories from 20 to 21 November.

  1. ExposedA medical imaging annotation server is reachable from the internet
  2. MapAI agents enumerate its API without a human driving each step
  3. FindThey identify a path traversal flaw in file upload
  4. ExploitArbitrary file writes on the server follow
  5. EscalateReported: a Python start-up hook leads to root
  6. CaughtThe attackers left their own agent console exposed
The MONAI Label case as reported: a research tool, an exposed server and agents doing the discovery.

1. AI agents find and exploit a medical imaging zero-day

Amazon Threat Intelligence reported that a suspected China-nexus operator used autonomous agents from the open-source Cairn framework to map the API of MONAI Label, an open-source medical imaging annotation tool, find a path traversal flaw in a file upload function and exploit it against university servers in East Asia1. The Register reports the agents went on to gain root by planting a Python start-up hook, and that Amazon found the campaign because the attackers left their agent instance publicly reachable2.

CTO take: this case stands out because the agents found a new flaw rather than only chaining known ones, and the target is telling: a research tool deployed by a lab, exposed to the internet, outside normal IT oversight. Hospitals, universities and health-tech companies should find every research and annotation server that is internet-facing and put it behind authentication and a VPN. Note the escalation route too: the same Python start-up hook technique as the agent-framework compromise. See AI-assisted intrusions.

2. An Entra Connect Health flaw exposes tenant metadata

Microsoft disclosed that attackers exploited an API flaw in Entra Connect Health to read tenant-level metadata, including administrator email addresses and tenant configuration; it fixed the flaw on 19 November, is notifying affected customers and published a read-only script so administrators can see what may have been exposed3. Reports say passwords, hashes and tokens were not exposed.

CTO take: leaked administrator emails plus tenant details are a ready-made target list for phishing and impersonation, so the risk outlives the fix. Check Azure Service Health for a notice, run Microsoft's script, warn your administrators specifically, and confirm every admin account uses phishing-resistant sign-in. Consider separate, unpublished addresses for administrator accounts. See the access hardening case study.

3. House committees recommend limits on Chinese open-weight models

Two House committees published a joint report on Chinese open-weight models such as DeepSeek, Qwen and Kimi, recommending restrictions on their use in federal agencies and contract work, a national American open-weight model initiative and a model provenance standard4. CNBC reports Cursor told the committees it would explore phasing out Chinese-developed open-weight models because of geopolitical risk5.

CTO take: recommendations are not rules, but US companies that sell to government, or to customers who do, should expect questions about where their models come from. Know which models, including fine-tunes and distilled derivatives, sit inside your products and tools, and keep an abstraction that lets you switch. That is good practice regardless of politics. See per-market AI models, model routing and my US page.

4. Signal blocks operating system AI assistants from reading chats

Signal now blocks screenshots and screen recording by default on supported iPhones and uses its screen security feature on Android, to stop Apple's and Google's AI assistants from reading chats on screen; it says it cannot do the same on macOS and has asked platform makers for a proper way to keep AI agents out of sensitive apps6.

CTO take: operating system assistants that read the screen are a new data path for every app that shows sensitive information: banking, health, internal tools. Product teams should check what their apps expose to screen-reading assistants, use the platform flags available for sensitive screens, and update privacy notices. Security teams should add OS assistants to their data flow maps. See on-device data and consent.

What I'd do this week

  • Find internet-facing research and annotation servers and put them behind authentication.
  • Run Microsoft's Entra Connect Health script and brief administrators about targeted phishing.
  • List every model, including derivatives, in your products and tools, with its origin.
  • Check which sensitive screens in your apps are visible to OS AI assistants.

I am a fractional CTO and AI architect based in India, the United States and the UAE, working globally. If any of these lands on your desk this week, analyse your challenge and bring the result to a call.

References
  1. 1
  2. 2
  3. 3
  4. 4
  5. 5
  6. 6
FAQ

Common questions

What did Amazon report about AI agents?

Amazon Threat Intelligence said a suspected China-nexus operator used autonomous agents from the open-source Cairn framework to map MONAI Label's API, find a path traversal flaw in a file upload function and exploit it to write files on a university server.

What did the Entra Connect Health flaw expose?

Microsoft said attackers used an API flaw to read tenant-level metadata, including administrator email addresses and tenant configuration. Reports say passwords, hashes and tokens were not exposed. Microsoft fixed it on 19 November and is notifying affected customers.

Is the House report on Chinese models law?

No. It is a joint committee report with recommendations, including restricting Chinese-developed models in federal work and creating a model provenance standard. It does not change rules by itself.

Keep reading

Related

ServiceFractional CTO and AI architect for US companiesSenior technology leadership for US founders, enterprises and investors: AI that reaches production, data platforms that pass security review, and governance that keeps pace with federal and state rules.→ServiceAI & LLM architectureI design and build AI systems that leadership can trust: grounded in your data, measured by evaluations, governed by your rules and priced so they make sense at scale.→InsightTech news, 18 November 2026: the Optimizely supply chain attack, Nvidia's first results without Huang, New York's AI enforcement unit and Microsoft's MAI Code 2Attackers used a flaw in Optimizely's experimentation platform to inject malicious JavaScript into the websites of 178 organisations, a reminder that every third-party script runs with your site's full trust. Nvidia reported strong results and raised guidance in its first quarter under an interim chief executive, New York set up a unit to enforce its AI safety law, and Microsoft released cheaper coding models.→InsightAI inference providers are now supply chain: what the NEAR AI and Fireworks breaches teach buyersWithin days of each other, NEAR AI and Fireworks AI disclosed breaches. One came through a flaw in an open-source inference engine running inside a confidential computing enclave; the other through a cloud provider hosting research machines, reaching customers' secrets, datasets and partial model weights. Most companies now send prompts, documents and fine-tuning data to providers like these. Here is what to learn, and what to ask.→
Open to advisory, fractional CTO & select builds

Let's talk about what you're building.

The fastest route is a free 30-minute call. Prefer to write? Send a message and I'll reply personally within two business days.

Start a conversation

How would you like to reach me?

Book a 30-min callFastest. Pick any open slot on my calendar.→ Write in GmailOpens a pre-filled draft in your browser.→ Use my email appOutlook, Apple Mail or your default client.→

I reply personally within two business days. Time zones from the Americas and Europe to the Middle East and Asia-Pacific.