All Posts
Anomali Cyber Watch
1
min read

Anomali Cyber Watch: SCKit MSP Worm & GhostAction GitHub Theft

Published on
November 17, 2026
Table of Contents

At a Glance

Attack Patterns
58
Exploit Public-Facing Application
3
Command and Scripting Interpreter: Powershell
2
Boot or Logon Autostart Execution: Registry Run Keys / Startup Folder
2
Proxy: Internal Proxy
2
Ingress Tool Transfer
2
* Frequency counts reflect mentions across collected reports
Story #1  |  November 12, 2026

New worm-like ‘SCKit’ malware targets MSPs in ScreenConnect attacks

▶ expand
Threat actors exploited an authentication flaw in RMM+, a ConnectWise ScreenConnect extension, to gain unauthorized access to cloud-hosted ScreenConnect management panels. Versions 1.0.31 and 1.0.32 of RMM+ accepted a blank access key if an administrator had never configured one, allowing commands to execute without user login or multifactor authentication. Attackers used the platform’s Run Command feature to push commands as SYSTEM and deploy SCKit, a custom 16-file toolkit, across at least 288 endpoints at 48 organizations within 90 minutes. SCKit installed attacker-controlled remote access clients, stopped the legitimate ScreenConnect client, and deployed a component called RelayTrap. RelayTrap exploits CVE-2026-84869 (Common Vulnerability Scoring System score 9.9), a client-side missing authorization flaw in the ScreenConnect technician client that allows files to be transferred and executed through an active session without host confirmation. When a technician connects to an infected endpoint, RelayTrap attempts to steal the technician’s panel session cookie and replay it against the panel to push payloads through the Shared Toolbox to every device that technician manages. Researchers found this propagation stage failed in every observed case. ConnectWise patched CVE-2026-84869 in ScreenConnect 26.6.5 in September 2026, confirmed the patch held in testing, and subsequently disabled RMM+. No attribution has been reported.
Analyst Comment
The propagation stage failed in every observed case, but the precise reason matters: the patch held where it was deployed. ScreenConnect 26.6.5 was available two months before this campaign, and the exploit did not succeed against patched clients in any recovered log. Organizations that had not applied the September update may have faced a different outcome, so “failed” should be read as “blocked by the patch” rather than evidence the technique is unreliable. The initial access vector deserves equal weight. The RMM+ flaw was not a zero-day but a software default that accepted a blank access key if an administrator had never configured one, a configuration gap in an extension capable of executing commands as SYSTEM across an entire managed fleet. For MSPs, the immediate actions are concrete: confirm ScreenConnect 26.6.5 is applied, verify RMM+ is disabled, audit logs for Run Command activity without corresponding user login events, and review technician workstation hardening, since SCKit was designed to harvest session cookies from connecting technicians rather than from managed endpoints. For organizations that use an MSP, the exposure in this campaign came through the MSP’s management panel, not from activity on their own networks. The 288 endpoints reached within 90 minutes sits well inside most detection windows: Mandiant’s M-Trends 2026 report places the global median dwell time at 14 days, meaning the full campaign completed before most organizations would have known it had begun. The realistic ask for downstream organizations is not technical: it is whether existing contracts or oversight processes provide visibility into the MSP’s patch posture and RMM configuration, and if not, how to get it.
Story #2  |  November 9, 2026

ARTEX Goes from Pentesting Tool to Attacker Staple in Three Months

▶ expand
ARTEX is an open-source framework, published on GitHub on July 26, 2026, that uses coordinated AI agents to run penetration tests with minimal human involvement: an operator sets an objective and the agents plan and carry out the attack steps autonomously. Within weeks of publication, threat actors began using it offensively. That risk became concrete on October 25, 2026, when an AI agent running ARTEX began accessing a WorkOS internal tool outside its testing environment, completing a bulk extraction of business contact details, account histories, customer communications, and operational metadata on October 30; four control failures contributed: public network exposure, a predictable development credential, absent rate limiting, and insufficient monitoring. The researcher responsible came forward and confirmed the data was deleted. Separately, Storm-0940, a China-aligned group, exploited compromised ConnectWise application credentials to attempt sign-ins across more than 2,250 Microsoft 365 tenants at 2,900 attempts per minute; VAPOR PANDA, a second China-nexus group, targeted more than 200 organizations across government, defense, pharmaceutical, semiconductor, nuclear, and telecommunications sectors over 11 days, with post-exploitation confirmed at nine. ARTEX does not introduce new attack techniques; it runs established ones at machine speed, and Google researchers characterize current use as unsophisticated but effective.
Analyst Comment
The more significant detail here is not which tool attackers are using but how adoption became irreversible: more than 12,000 forks were distributed before the developer could contain it, making further proliferation structurally inevitable once the repository was pulled. The independent adoption of the same framework by at least five unrelated actor clusters, spanning China-nexus espionage groups and financially motivated criminal operators, is an informal vetting signal; it suggests ARTEX performed reliably enough that separate actors reached for it without coordination, and broader uptake among threat actors is a reasonable expectation rather than speculation. Defenders should resist the AI framing as a distraction. The WorkOS post-mortem is a checklist of unresolved fundamentals: an internet-exposed internal API, a predictable credential, no rate limiting, and a five-day detection gap. The access and extraction are the intelligence-relevant events; the researcher’s confirmation of deletion, while cooperative, does not change what was exposed or for how long. The current generation of these attacks succeeds primarily because of unresolved hygiene gaps rather than technical sophistication that outpaces the defender. For organizations with mature behavioral telemetry, hunting for a single identity authenticating across multiple service principals or requesting tokens for several cloud services in a short window offers a practical detection approach. For organizations without that telemetry capability, the WorkOS failure modes are the more immediately actionable starting point.
Story #3  |  November 12, 2026

Nightmare Eclipse Drops Windows Zero-Day Exploit ‘SignDrop’

▶ expand
Security researcher Abdelhamid Naceri, also known as Nightmare Eclipse, published a proof-of-concept exploit named SignDrop on GitHub on November 10, 2026, targeting an unpatched local privilege escalation vulnerability in the Windows Common Log File System (CLFS) driver. The flaw originates in a security mitigation Microsoft itself introduced on October 28, 2025, which was designed to stop attackers from exploiting CLFS by requiring log files to carry hash-based message authentication codes (HMACs), stored in companion .cnpf files alongside each log. Naceri found that this signing mechanism can be turned against itself: by preparing a specially crafted log file and coercing Windows Defender into opening it through a legitimate RPC call, he causes CLFS to generate a missing .cnpf file under Defender’s own SYSTEM-level process. Because the file creation appears to originate from the antimalware engine’s own trusted process, kernel-level tamper protection cannot distinguish it from normal activity, allowing an attacker to escalate from a standard user account to full SYSTEM privileges and potentially disable any antivirus or EDR product on the machine. Systems where the CLFS mitigation is active, indicated by a registry value of 0 or 1 at HKLM\SYSTEM\CurrentControlSet\Services\CLFS\Authentication\Mode, are vulnerable, a condition Naceri estimates affects approximately 80% of Windows installations. No CVE has been assigned, no patch is available, and disabling the mitigation reintroduces the original class of CLFS memory corruption vulnerabilities it was designed to prevent.
Analyst Comment
SignDrop requires local access to execute and should be understood as a post-exploitation tool rather than an initial access vector; its significance lies in what it does once an attacker has a foothold, not how they obtained one. At that stage, it offers a demonstrated path to SYSTEM privileges through a mechanism that directly undermines the trust model underpinning EDR tamper protection. Because the exploit operates through the antimalware engine’s own process, the file creation event it generates is indistinguishable from legitimate activity at the kernel level. Defenders should not assume their endpoint protection provides a reliable detection backstop against this specific technique. Practically, options are limited: organizations can confirm their exposure state via the registry key documented in the summary, but neither remediation path is clean. Disabling the CLFS authentication mitigation removes the SignDrop attack surface and reintroduces the memory corruption class of vulnerabilities the mitigation was designed to close. The appropriate response is to await a Microsoft patch, review whether existing telemetry is positioned to detect post-exploitation behavior rather than the exploit itself, and treat endpoint protection health as a priority rather than an assumption. The broader intelligence value here outlasts this specific vulnerability: a mitigation that, according to the researcher, sharply reduced exploitation of a component was itself introduced as an attack surface. That pattern merits close attention. Security controls are not exempt from the scrutiny applied to the components they protect.
Story #4  |  November 9, 2026

GhostAction Campaign Targets GitHub Commit History in Expanded Credential Theft Operation

▶ expand
A new variant of GhostAction, a supply chain campaign first documented in September 2025, used compromised GitHub maintainer accounts to push a credential-stealing workflow to 346 repositories on October 8, 2026, with the campaign reaching more than 30,000 repositories and approximately 1,900 compromised accounts by October 12. Named .github/workflows/security-audit.yml and committed under messages mimicking security tooling, the workflow abused GitHub Actions to collect credentials through two methods: targeting Actions secrets by name, identified from existing repository workflows before the payload was committed, and searching the working tree and reachable git history using 13 patterns targeting AWS credentials, Anthropic, OpenAI, and OpenRouter API keys, GitHub and GitLab personal access tokens, Google, Firebase, and GCP API keys, Slack tokens, and SendGrid credentials. Because credentials removed from current code can remain in reachable git history, the sweep can recover secrets maintainers may believe were already deleted. Collected data was sent over HTTP to 193.32.204[.]199. Affected repositories include Uber’s athenadriver and kitao/pyxel; the latter distributes through the Python Package Index (PyPI) and crates.io, with publishing credentials among those targeted. By October 9, 2026, successful executions were confirmed; no malicious releases had been observed on PyPI or crates.io at that time.
Analyst Comment
GhostAction’s history sweep highlights a credential exposure risk that persists even after code cleanup: removing a secret from the codebase does not remove it from accessible commit history, and any still-valid credential remaining in that history could be recovered and abused. Where historical secret scanning is in place, defenders should verify it extends across all branches and that detection coverage includes the specific credential types targeted here, as coverage varies by configuration and supported patterns. Any repository where a contributor account may have been compromised warrants a full history investigation alongside a review of the malicious workflow and its execution logs. The targeting of Anthropic, OpenAI, and OpenRouter API keys alongside AWS credentials is consistent with documented criminal demand for stolen AI service access, which can be resold or used to consume services at the legitimate holder’s expense. While the initial account compromise mechanism remains unknown, meaningful controls are still available: organizations should audit which third-party maintainer accounts hold write access to their repositories, enforce phishing-resistant authentication, restrict token permissions, and ensure credential rotation procedures account for secrets that existed only in historical commits.
Story #5  |  November 13, 2026

Chained PHP2Shell and Gaia OS Vulnerabilities Exploited in Check Point Gateway Attacks, Deploying TSGA Backdoor and ESXi Wiper

▶ expand
An unknown threat actor exploited Check Point Security Gateway Mobile Access portals via PHP2Shell (CVE-2026-109618 and CVE-2026-109619, CVSS v4.0 9.2), enabling unauthenticated remote code execution, then escalated privileges to root through CVE-2026-123039 (CVSS v3.1 7.8), a Gaia OS flaw in which a world-writable socket allows any authenticated local user to run commands as root. Patches were available one week before exploitation began on October 30, 2026, targeting Israeli organizations. A broader opportunistic wave on November 2 extended to multiple countries and sectors. Post-exploitation activity included PHP web shells, the TSGA passive backdoor disguised as a legitimate Check Point process, and TSProxy, an injected library that conceals attacker traffic within existing appliance processes. Attackers deployed the gogo scanner, moved laterally to a VMware ESXi host via the vpxuser account, and executed destroy_v6, which attempts to overwrite files with zeros before deletion and drops politically themed DREAM_ANNEXATION.txt notes. TSGA and TSProxy are assessed to be PhiliKit variants, previously linked to China-nexus group Armory Castle (formerly UNC5221); analysts lack sufficient evidence for high-confidence attribution and have not ruled out false-flag or smokescreen scenarios. The U.S. Cybersecurity and Infrastructure Security Agency (CISA) added CVE-2026-123039 to its Known Exploited Vulnerabilities catalog on November 13, 2026.
Analyst Comment
PHP2Shell is a flaw in the PHP runtime itself, not in Check Point’s code, which means any organization running affected, unpatched PHP deployments faces exposure to the same underlying vulnerabilities regardless of firewall vendor. For Check Point operators, the more pressing concern is what TSGA does after gaining root: it can tamper with binary firewall logs, so defenders investigating a potential breach through standard log review may already be working with altered evidence. TSProxy compounds detection difficulty further by injecting into existing appliance processes and intercepting attacker connections, concealing malicious traffic within legitimate appliance activity and making it harder to distinguish through standard network monitoring. The combination of sophisticated persistence tooling and politically themed destructive activity within the first wave on October 30 raises questions about the attackers’ objectives, leading Check Point to consider scenarios including initial-access brokering, a false flag, and destructive activity used as a smokescreen. For VMware environments, the vpxuser lateral movement path demonstrates how a compromised perimeter device can become an internal pivot point to critical hypervisor infrastructure. Patch, then hunt: given TSGA’s log-tampering capability, remediation alone does not confirm a clean environment.
Story #6  |  November 12, 2026

AI Domain Typosquatting Delivers ClickFix Lures, Store-Signed Browser Malware, and Search-Hijacking Extensions

▶ expand
Researchers documented a delivery pattern named TypoPrompting involving 186 lookalike domains across 27 AI service brands, including Claude, ChatGPT, DeepSeek, and Perplexity, with a subset routing mistyped visitors through traffic distribution systems to malicious delivery chains. In one observed Windows sequence, the Claude lookalike cloude[.]ai passed visitors through a fingerprinting and proof-of-work gate before selectively serving a ClickFix lure disguised as a Cloudflare verification check; the page copied to the clipboard a command invoking curl.exe to download and execute W901.ps1, with the host 185.138.88[.]3 represented as the decimal 3112851459, removing the dotted IPv4 format that pattern-based filters detect. Eight AI typos directed visitors through fake security check pages to Net Guard, a Microsoft Store-distributed application classified as a malicious browser hijacker with capabilities to change the default search engine while the user is idle, execute remotely supplied JavaScript through Chrome DevTools, and modify browser profiles and extensions without user disclosure. A third route led users from the Perplexity lookalike preplexity[.]ai to a Chrome Web Store extension whose manifest replaced the default search provider without requiring declared permissions; the extension belonged to a 30-extension kit, with 13 removed and the remaining listings showing 972,000 declared users as of November 3, 2026.
Analyst Comment
The Microsoft Store distribution of Net Guard is the most significant takeaway. Microsoft signing provides a credible trust signal without guaranteeing benign application behavior: the Store-distributed package contained browser-control capabilities enabled by default, including search hijacking, remote JavaScript execution, and browser-profile modification, without informing the user during setup. The extension route presents a related gap: search-provider hijacking through a browser manifest requires no executable JavaScript and no declared permissions, making it unlikely to surface in permission-focused extension reviews. Both routes point to controls that evaluate behavior rather than trusting distribution provenance or signatures alone: managed access to approved AI services, enforced extension allowlisting, and application policies that consider package identity and behavior alongside signing status. For ClickFix detection, correlating suspicious browsing with a cmd.exe chain executing curl.exe and subsequently launching PowerShell, particularly scripts downloaded to temporary directories with hidden-window or execution-bypass flags, is the more reliable detection path than monitoring the Run dialog directly. The decimal host representation (3112851459 for 185.138.88[.]3) is a useful campaign-specific indicator; broader detection rules targeting alternative IP representations in download-and-execute chains will outlast this specific infrastructure.

FEATURED RESOURCES

November 17, 2026
Research
CISOs

Australia’s Threat Report: What It Means for CISOs and SOC Leaders

CISO role has shifted from overseeing an alert-centric security function to governing the enterprise decision layer on which resilience now depends. In a world defined by cloud complexity, identity sprawl, AI-enabled processes, and machine-speed threats, the traditional SOC model no longer provides enough speed, context, or cohesion.
Read More
November 17, 2026
Anomali Cyber Watch

Anomali Cyber Watch: SCKit MSP Worm & GhostAction GitHub Theft

Technical threat intel: Deep dive into worm-like SCKit malware targeting MSPs, GhostAction GitHub commit scraping, and weaponized ARTEX pentesting tools.
Read More
November 11, 2026
Operationalized Threat Intelligence
Cyber Threat Intelligence

AI Coding Tool Configuration Files Are a Credential Leak Channel Security Teams Have Not Closed

The secrets sprawl problem has a new accelerant: AI coding tools are creating local permission files that record developer commands verbatim (credentials included) and a design gap means the per-machine protection built into those tools does not carry over to the repositories teams share.
Read More
Explore All