Floating servers and cyan data tubes exposing zero-day cyber threats.

Daily Cybersecurity News – November 21, 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

Chinese hackers use Cairn AI to find zero-day

High

What happened

Amazon Threat Intelligence found a publicly hosted instance of the open-source Cairn autonomous AI penetration-testing framework whose operators, writing prompts in Simplified Chinese with China Standard Time stamps, used it to attack East Asian universities.

Given target IPs and the goal of obtaining shell access, Cairn's agents planned and executed multi-hour campaigns. One failed after 148 actions over 16 hours (including a Log4Shell try); another burned six-plus hours on 30,000 credential combinations. The sole success came when agents found and exploited a previously unpatched path-traversal flaw in an open-source medical imaging AI application that Amazon assesses with high confidence was a zero-day. The same flaw was then used against a second university. Maintainers were notified and issued a fix; AWS and a government partner alerted the victims.

Who is affected

Universities in East Asia (specific institutions and country not named by Amazon). Organizations running the affected open-source medical imaging AI application before the patch.

More broadly, any environment where operators can spin up public or poorly secured instances of AI-driven offensive frameworks like Cairn.

Why it matters

AI agents can discover and weaponize unknown flaws in minutes and run continuously without human fatigue, outpacing many SOC triage cycles. This demonstrates how nation-state-aligned actors are already operationalizing open-source AI red-team tools for persistent university targeting and zero-day discovery.

The speed and autonomy raise the bar for detection and response; traditional alert handling may lag machine-speed exploitation.

How it could have been prevented

Immediately inventory and patch any instances of the medical imaging application once the maintainer fix is available. Restrict public exposure of AI penetration-testing platforms; require authentication, network isolation, and monitoring of their operational logs and prompts.

Hunt for anomalous multi-hour automated scanning, credential stuffing, or shell-access attempts originating from AI agent frameworks. Segment university research and medical systems and apply least-privilege access. Notify and coordinate with national CERTs when similar activity is observed.

Relevant professional terms

Zero-day
A software vulnerability that is unknown to the vendor and has no available patch at the time it is first exploited.
Autonomous AI penetration-testing framework
A system of coordinated large-language-model agents that accept high-level objectives in natural language, then independently plan, execute, learn from failures, and chain offensive actions without step-by-step human direction.
Source: The Record

Armory Castle exploits VMware vCenter zero-day

Critical

Attack chain overview

  1. Initial access via BeyondTrust Remote Support zero-day (CVE-2026-102764)
  2. Move to internal Linux appliances lacking EDR
  3. Deploy BONESAW to exploit vCenter CVE-2026-123890, create SSO admin, copy directory DB
  4. Use SAWSNAKE to forge SAML tokens and bypass MFA
  5. Access ESXi, extract AD database and registry hives from Windows VMs
  6. Exfiltrate via the BeyondTrust appliance

What happened

Suspected China-nexus group Armory Castle (formerly UNC5221) has been exploiting VMware vCenter authentication-bypass CVE-2026-123890 (CVSS 9.8) since at least early September 2026, more than two months before Broadcom's November 19 patch in VMSA-2026-0009.

The flaw lets a remote unauthenticated attacker obtain an administrator session on vCenter. Mandiant and Google Threat Intelligence Group investigated multiple incidents. Attackers typically entered via a BeyondTrust Remote Support zero-day (CVE-2026-102764), moved to unprotected internal Linux appliances, then used the custom Go tool BONESAW against vCenter to create an SSO admin account and exfiltrate the directory database. A follow-on tool, SAWSNAKE, forged SAML tokens to bypass MFA and impersonate users, after which they accessed ESXi hosts and pulled Active Directory data from Windows VMs.

CISA added CVE-2026-123890 to its Known Exploited Vulnerabilities catalog on 2026-11-20. Evidence indicates the group possessed the exploit by mid-July.

Who is affected

Any organization running unpatched VMware vCenter Server (versions prior to 9.1.1.0200, 9.0.2.0200, 8.0 U3m, or 5.2.5), including standalone, VMware Cloud Foundation, vSphere Foundation, Telco Cloud Platform, and Telco Cloud Infrastructure deployments, plus related SDDC Manager and VMware Tools for Windows.

Enterprises that also use BeyondTrust Remote Support are at elevated risk of the full attack chain.

Why it matters

vCenter is the central management plane for virtualized infrastructure. Compromising it grants control over virtual machines, enables credential and AD theft, MFA bypass, and persistent access deep inside the network. Espionage groups have chained it with edge-appliance zero-days for months of undetected access.

The long exploitation window before public disclosure and patch availability shows how high-value hypervisor management flaws become priority targets for sophisticated actors.

How it could have been prevented

Apply the Broadcom patches in VMSA-2026-0009 immediately for vCenter, SDDC Manager, and VMware Tools. There are no workarounds.

Hunt for indicators of BONESAW/SAWSNAKE activity, unexpected SSO administrator accounts, directory database copies, forged SAML tokens, and lateral movement from BeyondTrust appliances or EDR-less Linux systems. Review vCenter and ESXi logs for anomalous admin sessions. Isolate management interfaces and enforce network segmentation around virtualization infrastructure.

Relevant professional terms

Authentication bypass
A flaw that lets an attacker gain access to a system or privileged session without supplying valid credentials.
SAML token forgery
The creation of fraudulent Security Assertion Markup Language authentication tokens, often by stealing signing keys, so an attacker can impersonate any user and bypass multi-factor authentication.

North Korea group hides C2 on Solana blockchain

High

What happened

Cisco Talos reported that suspected North Korea-aligned group UAT-12643 (low-confidence assessment) is using the Overlord remote-access trojan whose encrypted command-and-control addresses are stored in memos on Solana blockchain transactions from a wallet the operators control.

The Go-based Overlord queries the wallet for updates, allowing C2 changes without malware redeployment. It resists removal: termination attempts trigger a Windows Blue Screen of Death, and it persists inside the Windows Recovery Environment so it survives full system resets. The toolkit also includes the Good Game Python infostealer and a cryptocurrency clipboard hijacker. Talos identified 35 commands covering keylogging, screenshots, webcam/audio, file/process control, scripting, lateral movement, and in-memory plugins.

An operational mistake left the actors' file server publicly accessible, exposing tools and configs. Lures targeted crypto holders, gamers, and internet-phone users.

Who is affected

Primarily cryptocurrency holders, gaming-account owners, and users of internet phone software (lures included fake financial statements, tampered Viber installers, and app-named files). Keyword searches in English, French, Spanish, and Portuguese suggest possible European victims.

Any Windows user who executes the malicious payloads; no specific victim count was published.

Why it matters

Blockchain-hosted C2 is hard to sinkhole or block with traditional network defenses because the ledger is public and decentralized. Persistence in the Recovery Environment defeats standard wipe-and-reinstall remediation.

Financially motivated DPRK-linked activity continues to blend stealthy C2, destructive anti-analysis, and direct crypto theft tooling, raising risk for individuals and organizations holding digital assets or gaming credentials.

How it could have been prevented

Monitor for processes that query Solana wallets or unusual blockchain memo traffic. Block known Overlord indicators and the public file-server infrastructure when available from Talos.

Ensure endpoint detection covers Recovery Environment modifications and BSOD-on-terminate behavior. Educate users on lures involving financial documents, chat apps, and gaming. Use hardware wallets and clipboard-protection tools for cryptocurrency; enforce application allow-listing where feasible. Reset or reimage systems only after verifying Recovery Environment integrity.

Relevant professional terms

Command-and-control (C2)
The channel and infrastructure attackers use to send instructions to malware and receive stolen data from compromised machines.
Blockchain-hosted C2
A technique that stores encrypted server addresses or instructions inside public blockchain transaction memos so malware can retrieve updated C2 locations without relying on traditional domains or IPs that defenders can easily block.

CISA adds four actively exploited critical flaws

Critical

The four KEV additions

CVEProductCVSSIssue summary
CVE-2026-129448Apache Roller9.8Unauthenticated RCE via XML-RPC deserialization
CVE-2026-76466Cisco Expressway9.8Hard-coded/persistent cookie path to root commands
CVE-2025-42884SAP NetWeaver EP6.5JNDI injection enabling info disclosure/modification
CVE-2026-123890VMware vCenter9.8Authentication bypass to administrator session

What happened

CISA added four vulnerabilities to its Known Exploited Vulnerabilities catalog on or around 20-21 November 2026, citing evidence of active exploitation: CVE-2026-129448 (Apache Roller deserialization, CVSS 9.8), CVE-2026-76466 (Cisco Expressway hard-coded credentials / persistent cookies enabling root command execution, CVSS 9.8), CVE-2025-42884 (SAP NetWeaver Enterprise Portal JNDI injection, CVSS 6.5), and CVE-2026-123890 (VMware vCenter authentication bypass, CVSS 9.8).

Apache Roller maintainers confirmed active exploitation of the XML-RPC deserialization issue. CVE-2026-76466 was exploited against a Russian organization after PoC release, enabling lateral movement. CVE-2026-123890 is tied to China-nexus Armory Castle (with BeyondTrust chaining). Public reporting on CVE-2025-42884 exploitation remains limited. Federal agencies face a 23 November 2026 remediation deadline under BOD 26-04.

Who is affected

Organizations running unpatched Apache Roller (4.0.0 through 6.1.4), Cisco Expressway Series (clustered devices), SAP NetWeaver Enterprise Portal, and VMware vCenter / related products (versions prior to the fixed releases listed in Broadcom's advisory).

Especially high risk for internet-exposed or poorly segmented management interfaces and for Federal Civilian Executive Branch agencies.

Why it matters

KEV listing signals confirmed real-world exploitation and triggers mandatory federal patching timelines. Three of the four score 9.8 and enable unauthenticated remote code execution or full admin takeover of core enterprise platforms (virtualization, collaboration, portals, content management).

Combined with espionage and opportunistic use, these create immediate high-impact exposure across government and private sectors.

How it could have been prevented

Patch immediately: Apache Roller to fixed November 2026 releases; Cisco Expressway per October 2026 updates; SAP NetWeaver per November 2025 fixes; VMware vCenter/SDDC/Tools via VMSA-2026-0009.

Prioritize internet-facing and management-plane instances. Hunt for exploitation artifacts (unexpected admin sessions, deserialization payloads, root commands on Expressway, JNDI anomalies, BONESAW-like tools). Segment affected systems, disable unnecessary XML-RPC or database services, and enforce strict access controls. Federal entities must meet the 23 November BOD deadline.

Relevant professional terms

KEV catalog
CISA's Known Exploited Vulnerabilities list of flaws that are confirmed to be actively used by attackers in the wild and that U.S. federal agencies must patch on accelerated timelines.
Deserialization of untrusted data
A vulnerability class in which an application reconstructs objects from attacker-controlled input without proper validation, often allowing remote code execution before authentication.

Signal stops AI agents screenshotting chats

Medium

How it works

  • Signal enables screenshot blocking by default when it detects OS AI agents attempting to read the chat screen on supported Android/iOS versions.
  • Builds on the earlier Windows 11 Screen security control introduced against Recall.
  • Users may toggle off the protection, but Signal warns this breaks its privacy guarantees for the user and contacts.
  • Limitation: does not block filesystem or other non-screenshot data access paths available to agents.

What happened

Signal announced it will block screenshots by default on Android and iOS to prevent built-in OS AI agents (agentic systems in iOS 27 and detectable Android assistants) from reading private chat contents displayed on screen.

The change extends earlier Windows 11 protections against Microsoft Recall. Signal Foundation president Meredith Whittaker called OS-level AI agents an existential threat because they scan screen content, capture Signal data, and make it available to local or cloud AI systems outside Signal's control, undermining end-to-end encryption guarantees for all conversation participants.

Signal acknowledges the protection is incomplete: it does not stop other access methods such as filesystem retrieval of already-present data. Users can disable the block under Settings > Privacy > App Security, but doing so exposes content to the agents.

Who is affected

Signal users on iOS 27+ and Android versions where the app can detect AI screen-reading assistants, plus Windows 11 users already covered by the prior setting. All participants in a conversation are potentially affected if any party's device allows agent access.

Why it matters

OS AI agents can silently exfiltrate encrypted-messaging content after it is decrypted for display, breaking the privacy model users rely on. Without robust app-level opt-outs from vendors, messaging and other sensitive apps lose the ability to enforce their own guarantees.

The issue is already practical: reports show agents bypassing official connectors to reach Signal data. Builders of privacy-critical software must treat screen and OS-agent access as a new attack surface.

Relevant professional terms

End-to-end encryption
A method that scrambles messages so only the intended sender and recipient can read them; even the service provider cannot access the content.
Agentic AI screen access
Operating-system AI assistants that continuously or on-demand capture and process whatever is rendered on the device display, including content from third-party apps, often sending it to local or remote models for summarization or action.

Transcription firm leaks police bodycam data

High

What happened

San Francisco transcription company Tigerfish, used by police, prosecutors, government departments, and tech firms, left a large volume of customer data on a publicly accessible Amazon S3 bucket with no password and an easily guessable address.

Exposed material included transcriptions of police interviews, body-cam footage, raw customer audio and video, and invoices identifying law-enforcement clients. Independent researcher Anurag Sen discovered the open bucket in October 2026, attempted to notify the company (email landed in spam), then alerted TechCrunch. After contact, Tigerfish blocked all public access and confirmed other buckets were locked down.

Who is affected

Police departments, prosecutors' offices, government clients, and Silicon Valley tech customers of Tigerfish whose interview transcripts, body-worn camera video, source media, and billing records were exposed.

Scale is described as large but exact file or record counts were not published; anyone whose sensitive audio/video was sent to the firm for transcription.

Why it matters

Body-cam and interview material often contains highly sensitive personal data, criminal-investigation details, and identities of officers, victims, and witnesses. Public exposure creates privacy, safety, legal, and evidentiary risks and can undermine ongoing cases.

The root cause was a basic cloud misconfiguration, illustrating how simple storage errors at specialized processors can affect many high-stakes organizations at once.

How it could have been prevented

Audit all cloud storage (S3 and equivalents) for public access, directory listing, and predictable URLs; enforce block-public-access account settings and least-privilege bucket policies.

Enable access logging, monitoring, and alerting for public ACL or policy changes. Require encryption at rest and in transit, short-lived signed URLs for legitimate sharing, and regular external penetration or configuration reviews. Maintain out-of-band notification channels so researcher reports are not lost to spam. Notify affected customers and regulators promptly once exposure is confirmed.

Relevant professional terms

Cloud storage misconfiguration
An error in permissions or settings on a cloud bucket or container that unintentionally makes private files readable by anyone on the internet.
S3 bucket
An Amazon Simple Storage Service container used to hold objects (files); if its access-control list or policy is set to public, the contents can be listed and downloaded without authentication.

40 credit unions hit by unnoticed supplier breach

High

What happened

Hackers exploited a vulnerability in systems of Belfast-based software supplier Wellington IT, a data processor for credit-union mobile banking apps, accessing payment details of members at more than 40 Irish credit unions.

The attack occurred in early October 2026. Wellington closed the unidentified flaw on 6 October but initially did not recognize a data-protection breach; credit unions were not notified until 16 November, a roughly six-week delay. Ireland's Data Protection Commission received over 40 breach notifications. The UK ICO began seeing related reports, indicating possible Northern Ireland or broader UK impact. Officials state there is no evidence of funds access or fraudulent transactions so far; only payee details of app users were affected.

Who is affected

Members of more than 40 Irish credit unions (and potentially UK ones) who used Wellington IT-operated mobile banking apps; specifically their stored payee/payment details. Non-app users were unaffected.

Wellington IT customers in Ireland and Northern Ireland; exact account counts still under external review.

Why it matters

Supply-chain compromise of a shared banking-app provider instantly multiplies impact across dozens of financial institutions. Delayed detection and notification left members and credit unions unable to monitor for secondary fraud or take protective steps for weeks.

Even without immediate fund theft, exposed payee data enables targeted social engineering and future account takeover attempts, and triggers cross-border regulatory scrutiny.

How it could have been prevented

Credit unions and similar institutions should demand timely breach-detection SLAs, independent audits, and contractual notification deadlines from critical processors. Monitor for anomalous access to payee or beneficiary data and enable enhanced transaction alerts for affected members.

Wellington IT and peers must improve vulnerability response, ensure forensic processes correctly identify personal-data involvement, and maintain tested out-of-band customer notification paths. Apply least-privilege and segmentation around payment-detail stores; conduct regular red-team exercises on supplier connections. Regulators and NCSC coordination should be engaged early.

Relevant professional terms

Data processor
A company that handles personal data on behalf of another organization (the controller) rather than deciding the purposes of the processing itself.
Supply-chain breach
A security incident that begins at a third-party vendor or service provider and cascades to that provider's customers, often exposing data or systems the customers themselves did not directly control.
Source: The Record