Weekly Cyber Roundup: November 2-8, 2026

Aquí están las noticias principales que te perdiste la semana pasada.

Key takeaways

• Zyxel: The exploit behind 996 compromised GS1900 switches still worked on switches that had the June patch. The real fix, firmware ending in .3C0, shipped October 22.
• WordPress: CISA says attackers are exploiting CVE-2026-119977 in the core updater. Update to 7.1.4 or your branch’s fix, and block web access to wp-content/upgrade/ first.
• Binance: The post-mortem traces the $6.34 billion theft to a tampered Homebrew install, a fake job interview, and an internal AI tool.
• cPGuard: A hijacked update channel put a root backdoor on about 2,950 web hosting servers in 70 minutes. Update to 5.88.26 or later and rotate every credential on affected servers.
• WorkOS: ARTEX, an open-source AI penetration testing agent, breached an internal sales application, and researchers have read the exposed servers of one ARTEX operator.

1. Zyxel: The exploit behind 996 compromised GS1900 switches still worked after the patch everyone was told to install

El titular: Zyxel patched CVE-2026-107441, a CVSS 8.8 out-of-bounds write in the web management interface of its GS1900 switches. Zyxel describes the attacker as LAN-based and unauthenticated, and a crafted HTTP request is enough to run operating system commands. GreyNoise reported that a Chinese-speaking actor compromised 996 switches across 48 countries from around August 17 and took device configurations, network information, and hashed root credentials. Researchers tied those attacks to CVE-2026-7273, which Zyxel fixed in June and CISA added to its KEV catalog on September 21. Binarly then found the attacker’s exploit in an open directory on its command server, and the exploit still achieved unauthenticated command execution on all ten firmware releases that carried the June fix. That suggests the attackers were not relying only on CVE-2026-7273, and possibly not at all. Binarly reported its finding to Zyxel on September 29. Zyxel released firmware ending in .3C0 on October 22 and published its advisory on November 3. The flaw affects ten GS1900 models, including the GS1900-8, GS1900-24, and GS1900-48, on firmware ending in .2C0 or earlier. Help Net Security covered the disclosure.

Lo que estamos viendo en realidad

The patch defenders applied fixed a different bug. For six weeks the public record pointed at CVE-2026-7273, and a federal deadline set the benchmark for everyone else. Switches that installed the June fix stayed open to the exploit Binarly found. Zyxel shipped the fix for the real flaw 23 days after Binarly’s report and about 66 days after exploitation began, so a switch with every published patch applied was still exposed for more than two months. Binarly says advisories and version numbers cannot be blindly trusted, and it believes the exploit was LLM-generated.

The stolen material is the second problem. Configuration files and hashed root credentials give an intruder a map of the network and a head start on moving deeper, and the credentials stay valuable after the firmware is replaced. Restricting web management to a trusted network would have mattered under either CVE. Gestión de la Superficie de Ataque finds management interfaces reachable from the internet, and the same pattern appeared in our October 26 roundup, where Palo Alto said the risk was highest when the management interface was exposed. Binarly found this exploit because the attacker left a directory open, a mistake that also appears in story five.

2. WordPress: The core updater has carried a race condition since 2008, and the update that fixes it is itself exposed

El titular: CISA added CVE-2026-119977 to its Catálogo de vulnerabilidades explotadas conocidas on November 3, with a federal deadline of November 6. The flaw (CVSS 8.1) is a race condition in the WordPress core updater. Picus Security researcher Umut Bayram, who discovered it, explains the mechanics: during a core update, WordPress extracts the new package into a temporary directory under wp-content/upgrade, where a temporary installer finds no configuration file and starts setup. An unauthenticated attacker who reaches it in time points it at their own database, and the update then copies that configuration file over the site’s live wp-config.php. The attacker can create the first administrator and, on a typical single-site installation, upload a malicious plugin to run PHP on the server. WordPress fixed the flaw in version 7.1.4 on October 28 and backported the fix to every maintained branch down to 4.7. Versions 2.7 through 4.6 get no fix. ThreatDefence reports that Wordfence has blocked automated exploitation attempts since October 28 and that a working public proof of concept appeared on GitHub on November 4, and 2-Spyware quotes Wordfence saying threat actors are already actively attempting to exploit it.

Lo que estamos viendo en realidad

The update is the attack surface. A site is exposed during a core update, which is the one routine action administrators are told to take. The fix has a catch the headlines miss: the update that installs the fix is still performed by the old, vulnerable updater, so a site that clicks Update Now on an unpatched release leaves one last window open. Picus recommends blocking web access to wp-content/upgrade/ before updating, or updating manually outside the web root, and notes that the WP-CLI command wp core update uses the same updater.

The flaw sat in the updater for about 18 years. The researcher says he found it in roughly four hours using an AI model, and warns that the same speed is available to attackers. Exploitation attempts began six days before CISA’s listing, which is the gap between a patch and a public warning that most organizations treat as breathing room. Gestión de la Superficie de Ataque finds the WordPress installs that nobody remembers, including agency-managed and abandoned sites, which are the ones that stay on old versions longest.

3. Binance: The post-mortem traces the $6.34 billion theft to a fake job interview, a tampered Homebrew install, and an internal AI tool

El titular: Binance published its initial findings on November 3, describing a multi-stage campaign that it says was not a single breach or a simple phishing incident. It estimates about $6.34 billion in affected assets, with around $2.85 billion frozen or intercepted. In June, a Wallet DevOps engineer downloaded a tampered Homebrew installation while searching for a popular AI coding tool. Binance replaced the device and rotated credentials, but information about the engineer and his colleagues was already exposed. In September, someone posing as a recruiter approached another engineer through a legitimate job platform, ran two interviews, and asked him to download a technical assignment onto his work computer. The repository held malware that executed during setup and replaced a legitimate development tool component, and Binance says its endpoint tools did not detect it. On October 20, the attacker used that computer and an internal AI tool to take control of two more engineers’ machines and found a container deployment token on one of them. On October 29, stolen credentials gave access to another engineer’s computer, where the attacker deployed a container and ran a malicious script that reached the wallet signing service. The FBI has attributed the theft to North Korea’s TraderTraitor.

Lo que realmente estamos viendo: In our October 26 roundup we noted that Binance had not said how the laptop was reached. The answer is a job offer. Binance says its security team detected and prevented 57 suspected North Korea-linked cases in 2026, including fake job interviews, malicious developer tools, compromised packages, and fake video-conferencing applications, so this lure was not rare. Security Affairs notes that a September joint advisory on the WaterPlum group, also known as Contagious Interview, counted at least 30,000 infected devices in more than 100 countries. Iran-linked actors use the same lure: DarkBlinders sent engineers trojanized Visual Studio coding tests, as we covered in our DarkBlinders breakdown.

The internal AI tool is the detail to underline, and the least explained. Security Affairs notes that the report does not identify a flaw in an AI model. It shows an internal application becoming a pivot once an attacker held a compromised employee’s environment and the permissions that came with it. HackerNoon lists three open questions: how the tool was abused, which AI coding product was impersonated in June, and why a container deployment token sat on an engineer’s laptop.

Stopping withdrawals did not stop the theft. The attacker had used Ethereum’s EIP-7702 mechanism to give a malicious contract partial execution rights over a hot wallet, so the authorization lived on-chain independently of Binance’s withdrawal system, and with multi-party computation there was no single private key to revoke it. The last unauthorized transfer finalized at 23:56 UTC on October 29, and Binance cleared the authorization at 00:10 UTC on October 30. WaterPlum actors impersonate real companies, and Protección de marca detects and takes down the fake social media accounts and lookalike domains that impersonate your organization.

4. cPGuard: A hijacked update channel installed a root backdoor on about 2,950 web hosting servers in 70 minutes

El titular: OPSSHIELD disclosed on November 3 that an attacker used compromised credentials at 02:06 UTC on November 2 to reach its update distribution infrastructure and push malicious cPGuard versions 5.88.12 through 5.88.19. Reportes de Help Net Security that about 2,950 customer servers installed the fake update over roughly 70 minutes and ran it as root. The script planted a hidden root shell and collected password hashes, root SSH keys, the MySQL root password where present, root shell history, and cloud storage keys, which it sent to OPSSHIELD’s own licensing server under a compromised account. The attacker also took account records for about 7,900 customers, including bcrypt-hashed passwords, from the billing and licensing server. OPSSHIELD detected the attack within 12 hours, revoked the stolen update key and cloud credentials, and released a clean update, version 5.88.26, that removes the root shell.

Lo que estamos viendo en realidad

A security product became the delivery route, and it arrived with the privilege the product needs: root. cPGuard agents installed the fake update on their own and ran it as root, so 70 minutes was enough. OPSSHIELD’s 12-hour detection time is fast against IBM’s 2026 average of 247 days, and still slow against a script that only has to run once.

The way in was a chain, not a single password. According to OPSSHIELD’s own report, the attacker registered a customer account under a false identity the day before, installed a free trial to study how the agent talks to the licensing system, sent crafted requests to the billing portal’s callback pages, and gained command execution on the billing and licensing server. OPSSHIELD says it went from first probe to fake update in about ten hours and attributes that speed to attackers using AI tools, a claim we cannot verify. It has not yet determined how the attacker obtained the update key. Inteligencia de Credenciales detects stolen credentials across data breaches, paste sites, and dark web forums before attackers use them. Updating cPGuard does not undo the damage: the copied hashes and keys may still work, so every affected server needs new passwords and SSH keys. To check a server, look for versions 5.88.12 through 5.88.19 in /opt/cpguard/logs/install_log.

5. WorkOS: ARTEX, an open-source AI penetration testing agent, breached an internal sales application

El titular: WorkOS disclosed that on October 30 it identified unauthorized access to an internal application used by its sales team, exposing business contact records and summaries of customer communications. Founder Michael Grinich said the attacker used ARTEX, an AI penetration testing suite seen in multiple incidents this year, and cited the user agents artex-enrich/1.0 and norma/0.4 in the attacker’s web requests. ARTEX is an open-source framework developed in China. Seven South Korean financial institutions had data stolen between late September and early October, and CrowdStrike tied the activity to ARTEX, judging with moderate confidence that a lone, Chinese-speaking, financially motivated operator was behind it. ARTEX’s GitHub page and Docker image were taken down on October 11, but copies and forks remain online.

Lo que estamos viendo en realidad

The AI agent is the attacker’s labor, not a new kind of weapon. ARTEX automates reconnaissance and exploitation attempts, so every internet-reachable application gets probed at machine speed and the cost of looking has dropped close to zero. WorkOS lost an internal sales application, not production systems, and the South Korean attackers reportedly went after less protected systems used by employees, contractors, and loan agents rather than core banking networks.

The operators make the same mistakes as everyone else. In the Korean bank campaign the operator left servers open to the public, and researchers read the exposed files. Hunt.io’s analysis shows the operator started Claude Code 510 times since August 4, with a router sending every request through a reseller to DeepSeek and not to Anthropic. The Zyxel exploit in story one was found the same way, in an open directory on the attacker’s server, and DarkBlinders gave defenders a similar window: a GitHub token embedded in its loader opened its command repositories to Dream, as we covered in our DarkBlinders breakdown. South Korea’s regulator ordered banks to list every externally exposed system and close any path to internal data that skips authentication, which is sound instruction for any organization.

El patrón en las cinco historias

In every incident this week, the weak point was something the victim trusted: a patch, a routine update, a recruiter’s assignment, a vendor’s update channel, an internal application.

CybelAngel maps your external attack surface, including the suppliers, update channels, and management interfaces attackers scan for first. Our 2026 Supply Chain Cyber Risk Report shows how far that exposure reaches.