Isometric cyber machinery exposing AI agent and pentest tool threats.

Daily Cybersecurity News – November 16, 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

FBI flags threat actors abusing ARTEX AI pentest tool

High

What happened

The FBI issued a Flash notice on November 13 warning that multiple malicious actors have used the open-source AI-powered penetration testing framework ARTEX in global intrusion campaigns since at least August 2026.

ARTEX, released on GitHub in July 2026 by user Autumn-27, is written in Go with a Next.js console. It acts as an offensive C2 that orchestrates AI agents plugged into external LLMs such as Claude, GPT, DeepSeek, GLM, Grok and Kimi. Chinese-origin models, especially DeepSeek Flash, dominate observed attacker use due to cost and weak guardrails. Some campaigns reached privileged access in under three hours. The original repo was taken down in October 2026 amid abuse, but forks and a commercial ARTEX PRO version persist.

Documented impacts include theft of personal data from about 67,000 people at South Korean financial firms and a supply-chain compromise of ConnectWise Cloud Backup that reached more than 2,250 downstream tenants.

Who is affected

Organizations across financial services, government, technology, defense, semiconductor, energy, healthcare, manufacturing, telecommunications and legal sectors worldwide.

Any environment with internet-exposed or poorly segmented systems, partner portals, legacy services, CI/CD pipelines or cloud credentials is at elevated risk because the tool excels at chaining techniques against overlooked assets.

Why it matters

ARTEX lowers the skill barrier and multiplies campaign volume by automating enumeration and exploitation. Small teams can run parallel sophisticated attacks that previously required more time and expertise.

Its speed leaves defenders little window to respond, and its demonstrated supply-chain and trust-relationship pivoting capabilities turn weaker vendors into entry points for high-value targets.

How it could have been prevented

Harden and monitor overlooked systems such as partner portals, mobile backends, employee tools, legacy services and third-party integrations. Enforce least privilege, rotate credentials, restrict cloud access, and deploy network segmentation.

Monitor for noisy but rapid reconnaissance and exploitation patterns. Review LLM and AI infrastructure logs, application/API activity, and container/cloud indicators. Apply the FBI's recommended controls and report related artifacts to the Bureau.

Relevant professional terms

Penetration testing framework
A software toolkit that automates or assists ethical (or malicious) testing of systems for security weaknesses by chaining discovery, exploitation and post-exploitation steps.
Agentic AI offensive tooling
An autonomous multi-agent system in which a planner LLM decomposes goals and worker agents invoke tools and models to execute full attack chains with minimal human guidance.

AI agents target LiteLLM and Dify for API keys

High

What happened

Amazon Threat Intelligence captured records from decoy MadPot servers showing AI agents hunting internet-exposed AI applications such as LiteLLM, One API, Dify, Langflow, RAGFlow and others to steal stored AI provider API keys.

The agents verified keys, added them to shared pools, and used the stolen keys to power further searches. Amazon has moderate confidence the agents ran on ARTEX. Operators appear to be separate China-linked groups based on language and time-zone indicators. Playbooks listed default credentials, unfinished installs, auth bypasses, SSRF paths to cloud metadata, and secrets in JavaScript, source maps and environment files.

Cloudflare separately observed ARTEX activity after a compromised unrestricted CI/CD API token, including hundreds of database dump requests. Amazon could not fully verify all reported outcomes.

Who is affected

Operators of internet-exposed LLM gateways, proxies and AI app builders including LiteLLM, One API family products, Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat and Ollama, plus any organization storing multi-provider AI keys or using static CI/CD tokens.

AWS customers potentially affected were notified; broader impact includes any cloud or SaaS environment reachable via stolen keys or tokens.

Why it matters

Stolen AI keys enable both free inference for further attacks and direct abuse of paid model accounts. Focusing on multi-provider gateways maximizes blast radius from a single weak admin interface.

The self-reinforcing loop (keys fund more agent runs) and speed of agentic scanning mean basic hygiene failures are found and exploited far faster than traditional campaigns.

How it could have been prevented

Treat AI apps, gateways and inference endpoints as production systems. Change all default administrator credentials (including LiteLLM master keys), complete installs securely, and never leave management interfaces unauthenticated.

Scan configs, env vars, repos, containers and pipelines for secrets and rotate any that are exposed. Restrict API tokens with expiry, IP allow-lists and least privilege. Monitor for anomalous AI spend or usage and for ARTEX-related indicators such as specific Worker script names or known scanning IPs.

Relevant professional terms

API key
A secret token that authenticates and authorizes calls to a cloud or AI service so the provider can bill and control access.
SSRF (Server-Side Request Forgery)
An attack in which a vulnerable server is tricked into making HTTP requests to internal or cloud-metadata endpoints that the attacker could not reach directly.

BO Team uses Log4Shell to seize Russian code-signing server

Critical

What happened

Kaspersky ICS CERT investigated a months-long compromise of an unnamed Russian software developer's full IT infrastructure after the firm contacted them in June 2025. The intrusion is attributed to the Ukrainian-origin group BO Team (also known as Black Owl, Hoody Hyena, Lifting Zmiy).

Initial access came via an exposed SSH service on a Linux terminal server. Attackers then exploited CVE-2021-44228 (Log4Shell, CVSS 10.0, CISA KEV, actively exploited) on a VMware vCenter 7.0.3 server to gain root, deploy persistent backdoors (including hourly cron scripts pulling GSocket), and pivot across hypervisors, Windows and Linux hosts, GitLab, Ansible, domain controllers and ultimately the code-signing server.

They stole source code, private keys, credentials (via lsass dumps, ntds.dit, registry) and entrenched with LotL techniques, timestomping and log deletion. No ransomware or public claims occurred; the goal appeared to be supply-chain preparation. Access was lost after detection; a later phishing attempt using a stolen mail-server key was blocked.

Who is affected

The primary victim is a Russian software developer whose products serve a domestic customer base. All components of its IT estate were compromised, including development/CI/CD systems and the code-signing server.

Any organization still running unpatched Log4j-vulnerable VMware vCenter or similar Java services, or exposing SSH, faces similar risk from this and other actors.

Why it matters

Control of a code-signing server and CI/CD pipeline enables classic supply-chain attacks that can push malicious updates to every customer of the victim. Even without completing the chain, the theft of source and signing material creates long-term risk.

The case shows Log4Shell remains lethal years later when unpatched, and that determined actors will live off the land for months to reach high-value signing infrastructure.

How it could have been prevented

Patch or mitigate CVE-2021-44228 everywhere Log4j is present, especially VMware vCenter and other management planes. Immediately restrict or remove internet-facing SSH and enforce key-based auth with MFA where possible.

Segment development, signing and production networks; monitor for unusual cron jobs, GSocket or similar reverse tunnels, lsass/ntds access and anomalous domain-admin activity. Rotate all credentials and signing keys after any suspected compromise and hunt for persistence.

Relevant professional terms

Log4Shell
A critical remote-code-execution vulnerability in the Apache Log4j logging library that lets attackers run commands by sending specially crafted strings that the logger processes.
Code-signing server
A hardened system that holds private cryptographic keys used to digitally sign software binaries so operating systems and users trust them as authentic and untampered.

Stolen Fireworks AI key trojanizes Claude extension

High

What happened

Cursor disclosed that an attacker obtained an access key Cursor had issued to Fireworks AI. The key was exposed in a Fireworks breach and used to copy data from a Cursor read-only storage bucket containing coding-agent traces from non-Privacy-Mode sessions (mostly 2025 data allowed for training).

Those traces included unredacted credentials for less than 0.01% of customers. The attacker leveraged them to reach Cursor SaaS tools and exfiltrate limited source repos, support data, usage metrics, conversation summaries and some user-supplied credentials.

On October 29 2026 the attacker used a stolen internal credential to place trojanized builds of Anthropic's official Claude Code extension (versions 2.1.312 and 2.1.313) into the marketplace cache. The malicious versions attempted to install a persistent SSH backdoor. At most 5,000 users who requested those cached versions during an approximately 11-hour window were potentially affected before Cursor removed the builds, took the marketplace offline and sank the attacker's server.

Who is affected

Up to about 5,000 Cursor users who installed the specific malicious Claude Code extension versions from the marketplace on October 29. Customers whose non-Privacy-Mode session traces contained credentials, plus anyone whose data appeared in the limited support or metrics sets.

Cursor's production environment, Privacy-Mode users, other extensions and official app updates were not impacted. Fireworks AI customers including Uber and Notion were separately affected by the upstream breach.

Why it matters

A supply-chain style compromise of a popular AI coding extension can deliver backdoors directly into developer workstations, which often hold source code, cloud credentials and production access.

The incident illustrates how a single vendor key and residual training data can cascade into marketplace poisoning, and how quickly such an implant can be distributed before detection.

How it could have been prevented

Enable Privacy Mode in Cursor to keep session data out of training buckets. Rotate any credentials that may have been present in older traces or sent to support. Monitor endpoints for unexpected SSH backdoors or anomalous extension behavior.

Vendors should issue short-lived, tightly scoped keys to partners, avoid long-lived shared storage of sensitive traces, and implement integrity controls and rapid revocation for marketplace caches.

Relevant professional terms

Trojanized extension
A software add-on that looks like a legitimate update or plugin but secretly contains malicious code such as a backdoor.
Marketplace cache poisoning
An attack that inserts malicious packages into the temporary storage layer that serves extensions or packages so users receive the bad version when they request a specific build.

UnrealIRCd CVE-2026-114713 exploited in attacks

High

What happened

Threat actors began exploiting CVE-2026-114713 in the open-source IRC server UnrealIRCd one day after public disclosure. The high-severity flaw (CVSS 8.6) is an improper neutralization of delimiters in the labeled-response module.

It fails to escape the IRCv3 label message tag when echoing replies, allowing any user on a network with two or more linked servers to inject server-to-server commands that execute with full server/IRCOp authority. An ordinary user can therefore KILL, GLINE and perform other privileged actions.

Patches and a hot patch shipped in UnrealIRCd 6.2.7.1 on November 7 2026 (affecting 6.0.0 before that release). Maintainer Bram Matthys called it the most critical issue in 15 years. Exploitation was observed by November 8; CISA added it to the KEV catalog on November 13 with a three-day remediation window for federal agencies. UnrealIRCd 5 and earlier are end-of-life and unpatched.

Who is affected

Operators of UnrealIRCd 6.0.0 through 6.2.7.0, especially multi-server linked networks. Older 5.x and prior releases are also vulnerable and will not receive fixes.

Any IRC network that allows untrusted users and has linked servers is exposed until patched or mitigated.

Why it matters

Full server-authority command injection from a regular user account lets attackers disrupt or take over IRC networks, kick users, ban at will and potentially pivot further depending on the deployment.

Rapid weaponization and KEV listing mean unpatched instances are under active attack; IRC remains widely used for communities, bots and some operational channels.

How it could have been prevented

Upgrade immediately to UnrealIRCd 6.2.7.1 or apply the official hot patch. Deploy the spamfilter workaround if patching is delayed. Replace any end-of-life 5.x or earlier servers.

Restrict who can connect, monitor for anomalous S2S commands or privilege actions from non-opers, and isolate IRC infrastructure from higher-value networks.

Relevant professional terms

IRC (Internet Relay Chat)
A long-standing text-based chat protocol and server software still used by communities, open-source projects and some automated systems.
Server-to-server command injection
A flaw that lets an unprivileged client craft messages which linked servers interpret and execute as privileged inter-server directives.
Source: SecurityWeek

Hackers probe newly disclosed Copeland controller flaws

High

What happened

VulnCheck disclosed 14 vulnerabilities in Copeland (formerly Dixell) XWEB EVO refrigeration supervisory controllers. Each flaw can yield root on test systems. Exploitation attempts against two of them began less than two hours after CVE publication on November 12.

CVE-2026-124014 (CVSS 9.1) allows unauthenticated arbitrary file write to the system configuration. CVE-2026-124010 (CVSS 9.8) permits unauthenticated root command execution via a setuid CGI. Both affect older firmware only. Nine flaws hit pre-2021 firmware (unauth RCE, hardcoded admin password, login-page injection). Five affect latest firmware, including predictable daily admin/root passwords derived from the exposed MAC address plus date, plus an authenticated command injection. Copeland ended support for XWEB EVO on March 16 2026 and will not fix the newer issues; customers are directed to XWEB Pro.

Sensors saw 125 requests overnight to the config-write flaw from one IP and later probes to the RCE flaw. Nearly 2,000 XWEB EVO instances remain internet-exposed.

Who is affected

Owners and operators of Copeland/Dixell XWEB EVO refrigeration controllers, especially those still on older firmware or exposed to the internet. Roughly 2,000 such devices are reachable online versus only about 108 XWEB Pro units.

Facilities such as supermarkets, cold-storage and industrial sites that rely on these controllers for temperature management.

Why it matters

Root access on refrigeration controllers can alter set-points, trigger defrost cycles or disable alarms, risking food spoilage, safety incidents and physical disruption, as seen in the earlier French supermarket campaign.

End-of-life status plus rapid opportunistic scanning after disclosure means unpatched or internet-facing units are soft targets for both criminal and state actors.

How it could have been prevented

Immediately remove XWEB EVO devices from the public internet, place them behind VPN or strict network segmentation, and treat them as untrusted. Plan migration to supported XWEB Pro models.

Change any default or predictable credentials, disable unnecessary services such as SSH if possible, and monitor for configuration changes or unexpected command execution. Inventory all refrigeration controllers and apply compensating controls while replacing EOL hardware.

Relevant professional terms

ICS/OT controller
Industrial control or operational-technology hardware that directly monitors and commands physical processes such as refrigeration, manufacturing lines or building systems.
Unauthenticated root command execution
A vulnerability that lets a remote attacker run arbitrary operating-system commands with the highest privileges without supplying any username or password.
Source: SecurityWeek

Google Project Zero adopts 90-day disclosure

Medium

What to watch

  • Vendors accelerating patch development and alternative mitigations ahead of the 2027 and 2028 deadline drops
  • Increase in public technical details for still-unfixed bugs once the 90-day (then shorter) clocks expire
  • How enterprises adjust emergency patching and compensating-control playbooks for Project Zero-reported issues
  • Whether other major research groups adopt similar tighter Full-X style timelines

What happened

Google's Project Zero announced it will publicly disclose vulnerabilities it reports 90 days after notifying the vendor, whether or not a fix has shipped. The window shortens to 60 days on February 2 2027 and to 30 days on January 3 2028 under the new Full X policy.

This ends the previous 90 + 30 rule that allowed an extra month of soak time after a patch. Most grace periods and special cases are dropped. Exceptions remain for in-the-wild exploitation (handled under Google's broader policy), already-public bugs, and vendor-requested early disclosure. Hardware issues may still receive extra time on a case-by-case basis.

Project Zero cites three LLM-driven reasons: faster attacker discovery and exploitation, near-zero patch-to-exploit time, and a sharp rise in finding volume that makes negotiation unmanageable.

Who is affected

Software and hardware vendors that receive Project Zero reports, especially those with slower patch cycles or complex rollout processes. End users of products tested by Project Zero, who may see details of unfixed bugs published on a stricter schedule.

Security teams that track Project Zero advisories and coordinate enterprise patching.

Why it matters

Shorter, non-negotiable deadlines increase pressure on vendors to ship fixes faster or deploy non-patch mitigations, reducing the window in which only a few parties know about a bug.

At the same time, users and defenders must be prepared for technical details of still-unpatched high-impact bugs to appear publicly sooner, changing how risk is managed during the disclosure period.

Relevant professional terms

Vulnerability disclosure policy
The published rules a research team follows for how long it waits after telling a vendor about a bug before it makes the details public.
Patch gap
The interval between a vendor releasing a fix and the majority of users actually applying it, during which attackers can reverse-engineer the patch to target the unpatched population.