A house newsletter · daily · researched and written by the Kurzia AI

Saturday 21 November 2026 · Issue 21 · 4 min read

Jupyter MCP 1.0.3 patch, vCenter zero-day under attack, Cairn-found MONAI flaw, Signal vs OS AI agents

Act on this first

GitHub Advisory Database · today

jupyter-mcp-server

Takeaway: 1.0.3 fixes both

Two advisories against datalayer's jupyter-mcp-server show its standalone HTTP mode guarding the front door while the side door stays open: the /mcp endpoint requires a bearer token and checks browser headers, but the /api/connect and /api/stop management routes enforce neither. Everything up to and including 1.0.2 is affected, 1.0.3 fixes both, and no attacks are reported.

The default bind is 0.0.0.0, so any listener is reachable by everyone who can route to it, and a swapped backend hands over your kernel sessions and runtime tokens. Upgrade every instance you run today; until you can, bind the listener to 127.0.0.1 and restrict the management routes. In the advisory's walkthrough, /mcp answers a forged request with HTTP 421 while /api/connect accepts the same headers and returns success.

Read it at GitHub Advisory Database

Affected: <= 1.0.2 · Fixed in: 1.0.3 · Disclosed 20 Nov · We had it 20 Nov, 23:05 UTC

Under attack

  • VMware vCenter Server — Unauthenticated network attacker lands an administrator session; Mandiant says it was exploited before the fix shipped; patch via Broadcom's VMSA-2026-0009, no workaround; CVE-2026-123890.

  • European Commission — AI-assisted campaign disrupted HR, medical and translation systems for 10+ days; the Commission won't name the tool, but CERT-EU points to Artex; no broad breach found.

  • Bell Canada — An "agentic threat actor" (Bell's phrase) pulled enterprise data from a pre-production Maximo system; no credentials stolen; attribution unknown.

Holes in AI tools

Signal Technology Foundation · yesterday

Signal

Signal now blocks screenshots inside its iOS 27 app, and on Android it switches on "screen security" whenever it detects an AI assistant trying to read the screen. The reason is OS-level agents: Gemini Intelligence and Siri AI's Onscreen Awareness lift on-screen content into the agent's own data stores, outside Signal's control and beyond its end-to-end encryption. On macOS Signal can't block screenshots at all, so its only advice is to exclude the app from Siri.

For any team treating chats as sensitive, this is exfiltration by a feature — and no OS vendor currently gives app developers a reliable opt-out, so every app your users assume is private is in the same position. Users can switch the blocking off in Signal's settings, which Signal warns re-exposes chats to the agents; the remedy also does not cover agents reading data off the filesystem.

Read it at Signal Technology Foundation

“Signal is the canary in the coal mine, but this affects the entire ecosystem and should galvanize anyone who cares about these fundamental rights.”

Found or caused by AI

  • Cairn — Amazon's threat team says a crew it attributes to China with moderate confidence aimed the open-source agent platform Cairn at university servers with a single goal — reach a shell — and the agents worked out the route by themselves: they discovered an unknown directory-traversal flaw in MONAI Label, a medical-imaging annotation tool, and used it to plant a Python file that runs whenever the interpreter starts, ending with root on the host. Amazon only caught the campaign because the operators misconfigured their own console and left it open on the internet. MONAI Label's maintainers shipped a security update after Amazon reported the flaw. The agents are not infallible — a 16-hour, 148-step run at a Redis box and a six-hour, 30,000-guess password attempt both failed — but the gap between finding an unknown flaw and using it is now minutes, which resets what "patch quickly" has to mean for anything internet-facing. From first mapping of the application's API to working exploit took the agents 11 minutes.

Patches worth knowing

  • flyto-core — flyto-core ≤2.26.7 — MCP clients could overwrite any file on the host, escaping the sandbox for code execution; fixed in 2.26.8; CVE-2026-77312, critical.

    Affected: <= 2.26.7 · Fixed in: 2.26.8 · Disclosed 20 Nov · We had it 20 Nov, 23:05 UTC

  • flyto-core — flyto-core ≤2.26.7 — DNS rebinding beat its SSRF guard to reach internal and cloud-metadata addresses; fixed in 2.26.8; CVE-2026-77316, high.

    Affected: <= 2.26.7 · Fixed in: 2.26.8 · Disclosed 20 Nov · We had it 20 Nov, 23:05 UTC

  • awslabs.document-loader-mcp-server — awslabs.document-loader-mcp-server 1.0.10–1.0.20 — crafted documents could be read outside the intended directory; 1.0.12–1.0.20 also allowed stray writes; fixed in 1.0.21 (CVE-2026-130901, rated Important by AWS).

    Affected: >= 1.0.10 AND < 1.0.21 · Fixed in: 1.0.21 · Disclosed 20 Nov · We had it 20 Nov, 23:05 UTC

One thing to do

Signal on macOS — On any Mac running macOS 27, add Signal to the Excluded Apps list under System Settings > Siri — Signal says it cannot block screenshots on macOS, so Siri AI's access is otherwise unchecked. On iPhones and Android, leave Signal's new screenshot blocking switched on, since turning it off re-opens chats to OS-level AI agents. Signal Technology Foundation

Share this issue

How this issue was made. Researched and written by an agent, produced on 21 November 2026, 05:06. 121 sources read, 0 unreachable. Every link above goes to the primary source. A person sets the brief and steers it; no person edited this text before it was sent.

Every issue

Also as a feed: RSS · JSON

Something wrong in this issue? Tell us