1. Executive Brief
This week’s strongest pattern was the convergence of exposed infrastructure, stolen identity, and AI-enabled speed. CISA added vulnerabilities in security gateways, AI infrastructure, browsers, communications platforms, and remote-access products to its Known Exploited Vulnerabilities catalog. CISA KEV catalog
Reported incidents also showed attackers moving through trusted channels: an AI extension marketplace, cloud providers, support platforms, hardware supply chains, and employee identities. Cursor incident reporting TSMC incident reporting
The practical lesson is direct. AI governance cannot sit apart from identity, patching, software supply-chain security, and runtime authorization.
2. Threats and Active Exploitation
UnrealIRCd server-level command injection. CISA added CVE-2026-114713 to the KEV catalog on November 13, with a remediation deadline of November 16. The flaw allows a user on a network with at least two linked UnrealIRCd servers to inject server-to-server commands that execute with server authority. CISA KEV entry UnrealIRCd advisory
This is a general-security issue, but it illustrates a wider control-plane risk: a component that appears peripheral can gain authority through trusted internal relationships. Attackers were reportedly exploiting the flaw before its KEV listing. The Hacker News report
Check Point gateway exploitation. CISA added CVE-2026-123039 after identifying an incorrect permission assignment in Gaia OS DynamicCLI. An authenticated local user may reach a world-writable socket and execute commands as root. The affected products include Check Point Firewall, Security Gateway, Security Management Server, Multi-Domain Security Management Server, and related platforms. CISA KEV entry Check Point advisory
Reported exploitation of vulnerable Check Point gateways also involved PHP2Shell activity and deployment of a VMware ESXi wiper. Treat that incident as reported, and investigate for persistence, credential access, lateral movement, and destructive activity. BleepingComputer report
Agentic offensive tooling. Microsoft reported that the financially motivated actor Storm-3358 used the ARTEX agentic offensive framework against financial-sector organizations in October. Microsoft also reported that threat actors used ARTEX to search GitHub for exposed OAuth application secrets and then perform authenticated actions inside compromised tenants. Microsoft on ARTEX Microsoft on exposed secrets
This is AI-enabled, not evidence that AI independently caused the intrusions. The material risk is automation applied to weak identity controls, exposed secrets, and over-permissioned applications.
Identity theft and social engineering. The FBI attributed more than $400 million in cryptocurrency theft over 18 months to a DPRK-linked campaign using deepfake video calls, according to reporting on coordinated U.S. government action. The claim is reported through security coverage and should be validated against the relevant FBI release before being used for legal or regulatory reporting. The Hacker News report BleepingComputer report
3. Vulnerabilities and Patch Watch
Prioritize the following items based on CISA’s exploitation status and the stated deadlines:
- CVE-2026-116606 and CVE-2026-111560, Barracuda ESG and related products. The authentication bypass can allow unauthorized configuration changes. CISA listed the command-injection flaw as chainable with the bypass. The listed deadline for the two vulnerabilities was November 13. CISA KEV catalog Barracuda advisory
- CVE-2026-102895, GPUStack. The flaw can expose administrative access, model and worker secrets, API-key creation, and code execution through an inference-backend function. CISA’s listed deadline was November 12. This is AI-adjacent because it affects model-serving infrastructure directly. CISA KEV catalog GPUStack advisory
- CVE-2026-123158, OpenSSH on Linux. CISA describes an authentication-bypass issue and listed a November 12 deadline. Confirm affected distributions and package versions through the OpenSSH release information and distributor advisories. CISA KEV catalog OpenSSH release notes
- CVE-2026-124194, Microsoft Outlook. CISA listed a missing-user-interaction vulnerability that enables network spoofing, with a November 24 deadline. Apply Microsoft’s security update and review mail-client telemetry for suspicious activity. CISA KEV catalog Microsoft MSRC
- CVE-2026-20320, Cisco BroadWorks. The XML external entity flaw may expose sensitive configuration information to an unauthenticated remote attacker. CISA listed November 24 as the deadline. CISA KEV catalog Cisco advisory
- CVE-2026-125350, Chromium. A use-after-free flaw in the XML component may allow arbitrary code execution inside the sandbox through a crafted webpage. Patch Chromium-based browsers, including Chrome and Edge, through approved enterprise channels. CISA KEV catalog Google Chrome release
- PostgreSQL CVE-2026-6460. PostgreSQL released updates for 52 flaws, including a reported type-confusion issue that may let a database user who can create objects execute code with server privileges. The report states that C-language extensions must also be rebuilt for full protection. Treat this as reported until confirmed against the PostgreSQL security release. The Hacker News report
4. Breaches, Compromises, and Incidents
Confirmed by vendor. Cursor confirmed that attackers distributed a malicious Claude Code extension through its marketplace after a credential associated with Fireworks AI was exposed. Cursor said the stolen key was used to copy data from a read-only environment. Cursor incident reporting The Hacker News report
This is an AI-adjacent supply-chain compromise. The lesson is not limited to extensions. Model registries, coding agents, plugin stores, package repositories, and evaluation pipelines all need provenance checks, signing, scoped credentials, and rapid revocation.
Confirmed by Trezor. Trezor reported hidden implants in three Safe 5 wallets sold by the CryptoBilis reseller. The reported implants could capture recovery phrases and transmit them over mobile networks. BleepingComputer report The Hacker News report
Confirmed by Quantum Systems. The German drone company said it detected an incident in late October connected to an active cyberattack by hostile actors. The operational impact was reported as disruption to German military drone operations. BleepingComputer report
Reported, not independently verified here. TSMC confirmed that employee data had been leaked online after a threat actor offered 6.9 GB of allegedly stolen information for auction. BleepingComputer report
5. Current Security News
- Microsoft reported SCKit activity abusing ScreenConnect to deploy unauthorized remote access and target technician sessions. Audit remote-support tooling and technician identities. Microsoft Security Blog
- Microsoft reported GSocket deployment on compromised Linux servers to preserve encrypted remote access after an initial vulnerability is patched. Hunt for unexpected binaries, services, and outbound connections. Microsoft Security Blog
- Microsoft researchers reported model-supply-chain risks involving poisoned datasets and checkpoints that may trigger malicious behavior in coding agents. Microsoft Security Blog
- Microsoft Research described sleeper-agent experiments in which a small number of neuron edits caused malicious behavior in an agent model. This is research, not evidence of a live compromise. Microsoft Security Blog
- AMD SEV-SNP research described Mior and Skyfall attacks that may allow a malicious host to read protected VM data; Mior may also alter VM memory and forge an attestation report. The reported attack assumptions include hypervisor control, and Mior also requires BIOS control. The Hacker News report
- Italy fined Meta €805 million, according to reporting, over Facebook for Android accessing gallery content without explicit consent. The case reinforces the need for consent, data minimization, and permission review in AI-enabled consumer products. The Hacker News report
6. AI Security and Governance Implications
The week’s evidence supports four separate categories.
AI-enabled: ARTEX reportedly helped attackers find exposed OAuth secrets and automate authenticated activity. Deepfake video calls reportedly supported social engineering. These attacks increase speed and scale, but still depend on identity, access, and operational weaknesses.
AI-adjacent: GPUStack is part of the model-serving control plane. The Cursor extension incident affects AI development workflows. Poisoned checkpoints and coding-agent extensions create software-supply-chain risks before a model reaches production.
AI-caused: No item in the supplied evidence proves that an AI system independently caused a breach. Do not use that label for the reported ARTEX, deepfake, or agent incidents.
General security: UnrealIRCd, Check Point, Barracuda, Chromium, OpenSSH, BroadWorks, ransomware, GSocket, and hardware implants remain important even without an AI connection. CISA KEV catalog
Governance should therefore cover the full chain: model provenance, registry access, tool permissions, OAuth scopes, runtime authorization, secrets management, human approval for high-impact actions, and post-action logging. NIST’s AI Risk Management Framework, OWASP’s GenAI Security Project, CISA’s AI guidance, and the EU AI Act provide useful governance anchors. NIST AI RMF OWASP GenAI Security Project CISA Artificial Intelligence EU AI Act
7. What Security Teams Should Do This Week
- Reconcile internet-facing assets against the latest CISA KEV catalog. Escalate Barracuda, Check Point, GPUStack, OpenSSH, Chromium, Outlook, BroadWorks, and UnrealIRCd findings.
- Validate that patches changed the running version, not only the package repository state.
- Search for new local accounts, API keys, OAuth applications, and unexplained consent grants.
- Rotate exposed secrets and remove unused OAuth scopes. Require phishing-resistant MFA for administrators and developers.
- Hunt Linux systems for unexpected GSocket binaries, persistence, and encrypted outbound connections.
- Review ScreenConnect and other remote-support logs for technician-session abuse.
- Block unapproved AI extensions, plugins, model checkpoints, and registry downloads.
- Require signatures, provenance records, malware scanning, and approval before AI artifacts enter development or production.
- Add runtime controls for agents: least-privilege tools, destination allowlists, approval gates, rate limits, and full action logs.
- Test containment. Confirm that security teams can revoke an agent identity, disable a tool, quarantine a model, and rotate its credentials within minutes.
8. Source Index
Primary and official sources
- CISA Known Exploited Vulnerabilities catalog, updated November 13, 2026.
- Microsoft Security Blog, November 9-13, 2026.
- Microsoft Security Response Center, current advisories.
- Check Point security advisory, November 2026.
- Barracuda security advisory, November 2026.
- Cisco BroadWorks advisory, November 2026.
- Google Chrome release update, November 2026.
- NIST AI RMF, current framework.
- OWASP GenAI Security Project, current guidance.
- EU AI Act, current regulatory framework.
Reputable reporting
- The Hacker News, November 13-15, 2026.
- BleepingComputer, November 13-15, 2026.
- Dark Reading, November 12-13, 2026.
- KrebsOnSecurity, November 10-11, 2026.
9. AXEC Perspective
The operational boundary between AI security and enterprise security is disappearing. A compromised model registry, OAuth application, coding extension, or inference server can become an identity and supply-chain incident. Security leaders should govern AI systems as production systems with clear owners, narrow permissions, verifiable artifacts, strong telemetry, and tested containment.
Top comments (0)