Isometric cyan data pipelines exposing AI-driven zero-day exploits.

Daily Cybersecurity News – November 22, 2026

CyberRecaps is supported by its readers. We may earn an affiliate commission at no extra cost to you if you buy through a link on this page.

Ghost in the Wires by Kevin Mitnick

China Hackers Use AI Agents to Find MONAI Zero-Day

Critical

Attack chain highlights

  1. Operator sets goal of shell access and selects targets in Cairn.
  2. Agents map MONAI Label API and abuse unauthenticated upload for path traversal (approx. 11 minutes).
  3. Later session plants sitecustomize.py-style payload and triggers training to obtain root reverse shell.

What happened

Suspected China-linked operators used autonomous AI agents built on the open-source Cairn framework to discover and exploit a previously unknown path-traversal flaw in MONAI Label, an open-source medical imaging labeling server, gaining root on at least one university server.

The flaw, CVE-2026-133968 (CVSS 9.8), lets an unauthenticated attacker write files outside the intended storage directory via the image parameter on the PUT /datastore/image endpoint (and a similar label path). It affects MONAI Label 0.8.5 and earlier with authentication disabled by default. Amazon Threat Intelligence reconstructed the campaign from an exposed Cairn control panel: agents mapped the API and achieved arbitrary writes in roughly 11 minutes, later planting a Python startup file and triggering a training job for a root reverse shell. Exploitation began in September 2026. Maintainers merged a source fix on October 19 but no installable patched release exists yet.

Who is affected

Research teams and universities running MONAI Label 0.8.5 or earlier, especially default no-auth deployments reachable on the network. MONAI Label is used to annotate medical scans for AI model training and is distinct from the broader MONAI framework.

Scale is limited to exposed instances, but any writable code or config location under the service account can escalate to code execution. Amazon observed compromise of one university server and suspected a second.

Why it matters

Medical-imaging research infrastructure is high-value for espionage and can hold sensitive patient-derived data or model IP. Autonomous AI agents dramatically shorten the time from target selection to zero-day discovery and exploitation, lowering the skill bar for sophisticated actors.

Operators of scientific or healthcare AI tooling must assume unauthenticated services will be probed by automated agents, not just human attackers.

How it could have been prevented

Enable authentication immediately (set MONAI_LABEL_AUTH_ENABLE=True) and restrict the ANNOTATOR role. Isolate the service behind network controls so only trusted clients can reach it.

Run the process under a low-privilege account with minimal filesystem write rights. Monitor for unexpected file writes outside the datastore and for new Python startup hooks. Apply the merged source fix or upgrade once a release ships; rebuild from the fixed commit if needed.

Relevant professional terms

Path traversal
A flaw that lets an attacker supply ../ sequences or similar to write or read files outside the folder the application intended.
Autonomous AI agent
An LLM-driven system given a high-level goal that independently chooses tools, probes interfaces, and iterates toward exploitation without constant human step-by-step guidance.
High

What happened

Security firm pwn.ai demonstrated a now-patched sandbox-escape remote code execution flaw in Ghostscript that turns ordinary file uploads, previews, print jobs, and email attachments into command execution on the rendering server.

Tracked as CVE-2026-125396 (CVSS 7.8), the bug is a use-after-free in the Type 1 font loader (.putgstringcopy in gs_fonts.ps). A crafted PostScript or PDF triggers a type error, recovers the privileged .forceput operator, and disables -dSAFER checks. Ghostscript before 10.08.0 is affected. No known in-the-wild exploitation has been reported. Artifex fixed it on September 8 in 10.08.0. A related WordPress upload path was separately addressed as CVE-2026-65640 (CVSS 8.8).

Who is affected

Any Linux server or container that uses Ghostscript (directly or via ImageMagick, Pillow, libvips, Qt, etc.) to process untrusted PDFs, PostScript, EPS, or images that get handed off even when named .png/.jpg/.pdf.

Demonstrated paths include anonymous Debian print queues, Roundcube email attachment handling, public WordPress plugin upload forms (Forminator, Essential Addons for Elementor), and Author-level WordPress uploads. Bundled copies inside container images are often missed by OS updates.

Why it matters

Ghostscript is a silent dependency in countless document and image pipelines. A single crafted file can escape the intended sandbox and run as the renderer account, which is frequently enough for further compromise or data theft.

Because many apps invoke it indirectly, developers and operators may not realize they are exposed until a thumbnail or print feature is abused.

How it could have been prevented

Upgrade Ghostscript to 10.08.0 or later everywhere it runs, including inside container images and language-library bundles. Also apply the WordPress 7.0.4+ fix if using that platform.

Disable or sandbox unnecessary document conversion features. Run renderers under strict least-privilege accounts with no network and limited filesystem access. Prefer safer alternatives or content-type validation that rejects PostScript/EPS where possible. Monitor for unexpected process spawns from image/PDF libraries.

Relevant professional terms

Sandbox escape
Breaking out of the restricted environment an application deliberately places untrusted code or documents into so they cannot harm the rest of the system.
Use-after-free
A memory-safety error where code continues to use a pointer after the memory it points to has been freed, letting an attacker overwrite it with controlled data and often gain code execution.

Hackers Exploit VMware Cisco Apache Flaws in Espionage

Critical

KEV additions at a glance

CVEProductCVSSIssue
CVE-2026-129448Apache Roller9.8Unauth RCE via XML-RPC deserialization
CVE-2026-76466Cisco Expressway9.8Hard-coded creds to root commands
CVE-2025-42884SAP NetWeaver6.5JNDI injection info disclosure
CVE-2026-123890VMware vCenter9.8Authentication bypass

What happened

CISA added four vulnerabilities affecting Apache Roller, Cisco Expressway, SAP NetWeaver Enterprise Portal, and Broadcom VMware vCenter to its Known Exploited Vulnerabilities catalog on evidence of active exploitation.

The flaws are CVE-2026-129448 (Apache Roller deserialization RCE via XML-RPC, CVSS 9.8), CVE-2026-76466 (Cisco Expressway hard-coded credentials leading to root command execution, CVSS 9.8), CVE-2025-42884 (SAP NetWeaver JNDI injection for info disclosure/modification, CVSS 6.5), and CVE-2026-123890 (VMware vCenter authentication bypass, CVSS 9.8). All are confirmed in KEV. Apache Roller maintainers noted active exploitation. Cisco Expressway was used against a Russian organization after PoC release. VMware vCenter exploitation is linked to China-nexus group Armory Castle (formerly UNC5221), sometimes chained with BeyondTrust CVE-2026-102764.

Who is affected

Organizations running unpatched Apache Roller 4.0.0-6.1.4, Cisco Expressway Series (clustered), SAP NetWeaver Enterprise Portal, and VMware vCenter (specific 8.x/9.x and Telco Cloud versions listed in advisories).

Federal Civilian Executive Branch agencies face a November 23, 2026 patch deadline under BOD 26-04. Enterprise identity, collaboration, virtualization, and portal systems are primary targets.

Why it matters

These are high-impact, often unauthenticated or low-barrier flaws in core enterprise infrastructure. Active exploitation by espionage actors means rapid weaponization for persistent access, lateral movement, and data theft.

KEV listing forces prioritized remediation; failure leaves organizations open to nation-state level intrusion.

How it could have been prevented

Patch immediately to the fixed versions: Apache Roller (November 2026 release), Cisco Expressway (October 2026), SAP NetWeaver (November 2025 updates), and the listed VMware vCenter builds (e.g., 8.0 U3m, 9.0.2.0200, 9.1.1.0200).

Prioritize internet-facing and high-value instances. Monitor for anomalous XML-RPC, database-service, JNDI, and authentication-bypass activity. Apply network segmentation and least privilege around these services. Follow CISA KEV guidance and BOD deadlines.

Relevant professional terms

KEV catalog
CISA's public list of vulnerabilities known to be actively exploited in the wild, used to drive urgent patching priorities.
Deserialization of untrusted data
Converting attacker-controlled serialized objects back into live code or objects without proper validation, frequently resulting in remote code execution.

Microsoft Entra Connect Health Flaw Led to Tenant Data Theft

High

What happened

Microsoft disclosed that attackers exploited a vulnerability in the Entra Connect Health API to obtain unauthorized read-only access to tenant-level metadata, including administrator and technical-contact email addresses plus configuration details.

The flaw was introduced July 23, 2026, discovered November 9, and fixed server-side on November 19. Microsoft concluded its investigation November 20 and began notifying affected customers. No evidence of unauthorized configuration changes was found. Exposed data also included tenant IDs, display names, domains, country/region info, license tiers, sync status and timestamps, onboarding status, and an expired legacy certificate. Passwords, hashes, tokens, valid keys, AD topology, and subscription details were not exposed.

Who is affected

Organizations using Microsoft Entra Connect Health to monitor on-premises identity infrastructure. Microsoft has not published a total count of impacted tenants, though multiple administrators reported receiving notices.

Any tenant whose metadata was queried via the vulnerable API during the window is in scope.

Why it matters

Administrator email addresses and tenant configuration details are high-value for targeted phishing, social engineering, and reconnaissance that can precede deeper account takeovers or supply-chain attacks.

Even read-only cloud-service metadata leaks create lasting risk because the information cannot be fully revoked.

How it could have been prevented

No customer-side patch is required; Microsoft already deployed the server-side fix. Review Azure Service Health and any notification received.

Heighten monitoring for phishing that spoofs or targets the exposed admin/tech-contact addresses. Rotate any credentials or secrets that might have been referenced in related configurations as a precaution. Verify Connect Health agent registrations and monitoring settings for anomalies. Enable strong phishing-resistant MFA everywhere.

Relevant professional terms

Tenant metadata
High-level configuration and contact information about an organization's cloud directory that is not the actual user passwords or authentication secrets.
Entra Connect Health
Microsoft's cloud monitoring service that reports the health, performance, and alerts of on-premises identity components into the Entra admin center.

Malicious Page Hijacks Meta Muse Private Chats

High

Demonstrated attack sequence

  1. Malicious page loaded via Muse browser achieves code execution in agent container.
  2. Sentinel bypass lets conversation history and memories exfiltrate with no user prompt.
  3. Kernel exploit escalates to root on the per-user cloud VM.
  4. Root access reaches credential store, files, and Sentinel.

What happened

Researchers at NebuSec demonstrated that a single malicious webpage opened by Meta's Muse personal AI agent could exfiltrate the user's private conversation history and stored memories without any approval prompt, then escalate to full control of the user's isolated cloud VM.

The chain began with a flaw in Muse's built-in Chromium browser (used for tasks such as PDF generation), achieved code execution inside the agent container, bypassed the Sentinel egress monitor, and combined with a Linux kernel flaw to escape to root. From root the researchers reached the credential store for connected accounts, saved files, and Sentinel itself. The work was a laboratory demonstration only; Muse's team fixed all described issues before publication. No evidence of real-user exploitation was reported. The compromise stayed inside one user's VM.

Who is affected

Users of Meta Muse, the personal AI agent launched in September 2026 that runs each user's instance on its own cloud computer and can act on connected accounts.

Because the demo required the agent to process attacker-controlled content, any Muse user who lets the agent browse or generate from untrusted pages was theoretically exposed before the fixes.

Why it matters

AI agents that hold conversation memory, act on behalf of users, and store third-party credentials create a high-value single point of failure. Even well-designed isolation (per-user VMs, Sentinel gate, credential vaulting) can be defeated by chained browser and kernel bugs.

Builders of agentic systems must treat the agent itself as potentially hostile and continuously harden every layer the agent can reach.

How it could have been prevented

Muse users need take no action; the issues were fixed server-side by Meta before the research went public. As general hygiene, limit the accounts and permissions connected to any AI agent and review agent activity logs.

For operators building similar agents: keep embedded browsers fully patched and sandboxed, enforce strict egress allow-lists that cannot be bypassed by content type tricks, apply kernel hardening and rapid patching inside agent VMs, and assume prompt injection or malicious pages will occur.

Relevant professional terms

AI agent
Software that uses a large language model to plan and take real-world actions (browse, buy, message) on a user's behalf rather than only chatting.
Egress monitor
A security gate that inspects and can block or require approval for any data the agent tries to send out of its isolated environment.

Rackspace Spot Breach Exposes Kubernetes Secrets

High

Customer response checklist

  • Reissue Spot API tokens and replace automation refresh tokens
  • Download new kubeconfig files
  • Rotate all credentials inside cluster Secrets
  • Audit logs from November 17 onward for the listed IOCs

What happened

Rackspace notified customers that an unauthorized party used leaked credentials to access parts of its Rackspace Spot managed Kubernetes platform on November 17, 2026 (approximately 00:20-22:00 UTC).

Acting as legitimate users in some customer organizations, the attackers could create pods, list nodes and volumes, and read Kubernetes Secrets as well as Spot API tokens and refresh tokens. Rackspace learned of the intrusion on November 18 via a customer report, revoked sessions and tokens, removed unauthorized pods, rotated keys, and hardened controls. Indicators shared include IP 23.224.79.106, range 23.224.0.0/15, and pod/object prefixes i1, glm-, and probe.

Who is affected

Customers of Rackspace Spot whose organizations were reachable with the compromised credentials. Exact numbers have not been published.

Any Secret stored in affected clusters (passwords, tokens, API keys) and any Spot tokens held by those users were potentially readable.

Why it matters

Kubernetes Secrets are the de-facto vault for application credentials. Exposure can cascade into database breaches, cloud account takeovers, registry poisoning, and further lateral movement far beyond the original platform.

Credential-leak entry points remain one of the most reliable ways into cloud control planes.

How it could have been prevented

Immediately reissue Spot API tokens, replace automation refresh tokens, download fresh kubeconfig files, and rotate every credential stored in Kubernetes Secrets.

Review logs for databases, container registries, Git, cloud storage, and third-party services back to November 17 for abuse. Hunt for the published IOCs (IP 23.224.79.106, 23.224.0.0/15, prefixes i1/glm-/probe). Enforce short-lived tokens, workload identity where possible, and least-privilege RBAC going forward.

Relevant professional terms

Kubernetes Secret
A Kubernetes object that holds sensitive data such as passwords, tokens, or keys so pods can use them without hard-coding the values.
Refresh token
A longer-lived credential that an application can present to obtain new short-lived access tokens without forcing the user to log in again.

Signal Limits Screenshots to Block AI Scrapers

Low

How it works

Signal now blocks or warns on screenshot attempts in the mobile apps by default, matching the existing Windows behavior. The goal is to keep conversation contents from being silently captured by OS-integrated AI agents that scan the display. Users can disable the protection, but Signal then no longer guarantees privacy for those sessions. The change does not stop every possible AI or accessibility side channel.

What happened

Signal announced it is restricting users' ability to take screenshots inside its Android and iOS apps, extending a default block already present on the Windows desktop client.

President Meredith Whittaker framed the change as a response to the "existential threat" posed by operating-system-level agentic AI that can scan screen contents. Once captured by an OS AI agent, conversation data leaves Signal's control and can be stored or sent to local or remote models. Users may still disable the restriction, but doing so voids Signal's privacy guarantees for that session. Whittaker called the measure incomplete because other covert access paths remain.

Who is affected

All Signal mobile users on Android and iOS. The setting is on by default; users who need screenshots can turn it off at the cost of the app's privacy assurances.

Anyone relying on Signal for high-sensitivity conversations is the intended beneficiary.

Why it matters

End-to-end encryption protects data in transit and at rest inside the app, but cannot protect content once an OS-level AI agent or accessibility service captures the screen. As agentic features spread across Windows, Android, and Apple platforms, messaging apps face a new class of side-channel risk outside their traditional threat model.

Privacy tools must increasingly defend against the host operating system itself.

Relevant professional terms

End-to-end encryption
A system in which only the communicating users can read the messages; even the service operator cannot decrypt them.
Agentic AI
AI systems that autonomously observe the user's screen or apps and take actions, creating new data-exfiltration paths beyond classic malware.