
TL;DR
The 2026 cyber threat picture contains several genuine developments: AI-assisted intrusion frameworks conducting operational attacks, compromised cloud application credentials enabling access across customer tenants, exploited PHP vulnerabilities, malicious remote-management activity, third-party JavaScript injection, and attacks against industrial controllers. Investigations from Microsoft, Google Threat Intelligence Group, Unit 42, Huntress, government agencies, and affected vendors establish important facts. They do not establish that every incident was operated by the same actor, controlled by an AI agent, or part of one coordinated global campaign.
This tracker separates confirmed attacker operations from observed tooling, vulnerability exploitation, suspected attribution, and nonmalicious service disruption. It also distinguishes exposure from successful access, access from persistence, and persistence from demonstrated physical or business impact. The cases reveal recurring enterprise weaknesses, particularly excessive application authority, internet-exposed management paths, ungoverned executable dependencies, and insufficiently independent recovery controls.
The most useful connection between these incidents is architectural, not necessarily organizational. Attackers repeatedly exploit relationships that organizations have already authorized. The response is to map those relationships, preserve evidence of how they are used, constrain effective authority, and prove that containment removes access beyond the original entry point.
In This Article
- Introduction
- How to Read a Cyberattack Tracker Without Inventing Connections
- The 2026 Cyberattack Evidence Tracker
- Which Incidents Are Actually Connected?
- A Connection Map That Does Not Invent a Common Adversary
- Why the Reported Numbers Cannot Be Added Together
- What AI Involvement Can Actually Be Claimed?
- The Architectural Connection: Attackers Keep Crossing Established Trust Boundaries
- Turn the Tracker Into a Decision System
- A Practical Evidence Record
- How to Keep This Tracker Current
- External References
Introduction
A security leader receives several incident reports in the same week.
One describes a China-aligned actor using an AI penetration-testing framework to authenticate against thousands of customer tenants. Another documents malicious remote-access clients deployed through an MSP’s management software. A third identifies code injection affecting websites that had never changed their own application deployments.
At the same time, analysts are reporting autonomous exploitation frameworks, attacks on industrial controllers, and vulnerabilities requiring emergency patching.
The stories appear connected.
Several involve automation. Some involve the same vendor. Others target high-value infrastructure or describe persistent access that can survive the original compromise.
But similarity is not attribution.
Two incidents can use the same open-source tool without involving the same attacker. Two vulnerabilities can affect products from one supplier without sharing an entry point. An operational outage during a security update does not become a cyberattack simply because another organization is exploiting the vulnerability being patched.
Conflating these distinctions leads to poor decisions.
Security teams may hunt for the wrong indicators. Platform owners may assume that a vendor-side fix removes customer-level access. Executives may receive exaggerated victim counts. Recovery teams may declare an incident complete while a second credential or automation path remains active.
This evidence tracker takes a different role from the specialist articles in the series.
It is not another framework comparison, vulnerability guide, or incident-response runbook. It is an evidence register and an explanation of which reported activities are genuinely connected, which share only a technical pattern, and which should remain separate.
The tracker uses primary investigations and official advisories available through November 18, 2026. Historical incidents are labeled by their actual observation or publication period. The status of each record is a statement about the reviewed public evidence, not a guarantee that an incident has ended or that every affected organization has completed remediation.
Use the series overview to choose an entry point, the persistent-access guide to scope surviving authority, and the platform containment architecture and management-plane recovery guide to translate findings into recovery decisions.
How to Read a Cyberattack Tracker Without Inventing Connections
The first requirement is to classify what a source actually establishes.
A security advisory, a researcher-controlled malware sample, an observed authentication attempt, and a confirmed compromised customer environment are not equivalent evidence.
Six Evidence States
| Evidence state | What it establishes | What it does not establish |
|---|---|---|
| Confirmed exploitation | A credible investigation documents exploitation of a vulnerability or unauthorized action against an affected environment. | That every exposed system was compromised |
| Confirmed intrusion activity | Investigators observed unauthorized access, execution, persistence, or resource manipulation. | That the full attacker’s objectives succeeded |
| Observed attacker tooling | Researchers recovered or directly observed tools, infrastructure, or operational configurations. | That every advertised capability was successfully used |
| Targeting or attempted exploitation | Logs or investigations show a system, account, or organization was targeted. | Successful access or impact |
| Attributed campaign | A named research team connects activity to a tracked actor with a stated assessment. | A universally accepted legal identity or organizational relationship |
| Operational event, not established as an attack | A service failure or maintenance event is documented. | Malicious cause without separate evidence |
These states should remain distinguishable throughout an investigation.
For example, an unauthorized account added to a production system is evidence of persistent authority. It is not automatically evidence of scheduled destructive activity.
Similarly, an exploit framework recovered from an attacker staging server is important, but the capabilities present in its source code should not be reported as successful exploitation of every possible target.
Attribution Is a Separate Field
The incident record should capture the name used by the reporting organization and its stated confidence.
Microsoft’s Storm-0940 is a Microsoft tracking designation. See Microsoft’s campaign attribution.
Unit 42’s CL-STA-1238 is an assessment associated with the Cairn investigation.
Mandiant’s UNC5221 is associated with BRICKSTORM activity, but Mandiant explicitly distinguishes its tracking from another publicly named actor with which it has sometimes been conflated. See Mandiant’s actor-tracking distinction.
A shared country nexus does not prove that these groups are the same organization or that their operations are directed through one campaign.
One Incident Can Have Several Dates
Every maintained record should distinguish when activity occurred, when it was first observed, when it was disclosed, and when the source was last updated.
An investigation published in November may describe attacks from August or September.
A vulnerability disclosed in October may have been exploited before its official release.
A campaign discovered in 2025 may remain relevant to management-plane security in 2026 without becoming a newly observed 2026 campaign.
These dates support different questions, and they should not be combined into one generic “attack date.”
The 2026 Cyberattack Evidence Tracker
The following register covers the principal investigations examined throughout this series. It is a curated set of source-specific records, not an exhaustive global incident count.
Agentic Attacks, Cloud Trust, and Persistent Access
| Record | Activity and disclosure | Verified finding | Attribution and important limit |
|---|---|---|---|
| Storm-0940 / ARTEX / ConnectWise Cloud Backup | September–October 2026; Microsoft report October 21 | ARTEX accelerated authentication attempts against more than 2,250 customer tenants. Microsoft observed follow-on Graph and directory activity in some environments. | Microsoft attributes the campaign to China state-aligned Storm-0940. The tenant count is attempted reach, not confirmed breaches or exfiltration. |
| CAIRN / CL-STA-1238 | August–September 2026; Unit 42 report November 12 | Attackers conducted 25 operations against at least 19 organizations, achieving unauthorized access or retrieval in 10. Persistent administrator accounts and an API token were documented. | Unit 42 assesses China-nexus intelligence collection with moderate confidence. No evidence ties the operation to Storm-0940. |
exploit_agent | Identified in Google’s November 16 threat tracker | Researchers found an exposed attacker staging server containing an automated vulnerability-research and exploitation system, post-compromise tooling, and data associated with targeted victims. | Demonstrates observed offensive infrastructure and activity. The source does not establish a universal success rate or link the operator to all other agentic campaigns. |
| BRICKSTORM / UNC5221 | Historical context; Mandiant report September 24, 2025 | Long-term management-appliance backdoors, unauthorized accounts, infrastructure access, and sensitive VM cloning. Mandiant reported an average undetected dwell time of 393 days across investigated intrusions. | Suspected China-nexus espionage activity. Not reported as AI-operated malware and not a newly disclosed November 2026 attack. |
The strongest common pattern in these cases is the use of an initial capability to create or exercise additional authority.
Storm-0940 exploited preexisting application trust.
CAIRN-supported operations created persistent identities and credentials.
BRICKSTORM maintained long-term access through infrastructure appliances and management systems.
Those mechanisms belong to the same defensive problem class, but they are not evidence of one actor or campaign.
The exploit_agent investigation adds another dimension: a discovered system explicitly designed to automate patch analysis and exploitation. That is a meaningful escalation in attacker tooling, but it must be tracked separately from confirmed downstream effects.
Exploited Software, Remote Management, and Trusted JavaScript
| Record | Activity and disclosure | Verified finding | Attribution and important limit |
|---|---|---|---|
| PHP2Shell: CVE-2026-109618 and CVE-2026-109619 | Disclosed October 21; added to CISA’s Known Exploited Vulnerabilities catalog October 26 | Two critical PHP remote code execution vulnerabilities. Palo Alto Networks confirmed exploitation targeting exposed PAN-OS management interfaces. | Confirmed exploitation. The CVEs identify vulnerabilities, not one attacker or campaign. |
| ScreenConnect RMM+ / CVE-2026-127063 / SCKit | November 11–16, 2026 | Attackers used remote-management command capabilities to install unauthorized ScreenConnect clients and SCKit. Huntress observed 288 affected hosts across 48 organizations in one wave and 205 hosts across 41 in another. | RMM+ authentication weakness documented. The observed SCKit propagation chain from endpoint to technician to console was not confirmed successful. |
| ScreenConnect client / CVE-2026-84869 | Fix released September 8; successful exploitation documented November 12–13 | Huntress confirmed malicious file execution on a technician workstation through the vulnerable client. | Confirmed client-side exploitation. Not equivalent to successful execution of every guest-originated file request. |
| Optimizely Web Experimentation JavaScript injection | November 16; vendor report November 17 | Attackers exploited account-signup impersonation to modify Project JavaScript. Report URI observed 188 infected snippets associated with 178 organizations. | Confirmed upstream JavaScript modification and browser delivery. The figures do not measure compromised corporate networks or affected visitor counts. |
These records reveal different ways a trusted product can become an attacker-controlled execution path.
PHP2Shell involved vulnerabilities in software components.
RMM+ involved authentication within a third-party management extension.
CVE-2026-84869 involved authorization for file transfer and execution on a technician’s client.
Optimizely involved unauthorized changes to content delivered through a legitimate supplier integration.
The shared lesson is not that all software supply chains fail in the same way.
It is that organizations need to understand where executable authority resides, who controls it, and how an attacker could retain access after the initial problem has been addressed.
Industrial Control, Development Supply Chains, and Operational Disruption
| Record | Activity and disclosure | Verified finding | Attribution and important limit |
|---|---|---|---|
| Iranian-affiliated PLC activity | Activity since at least March 2026; joint U.S. advisory April 7 | Government agencies documented exploitation of internet-exposed PLCs, manipulation of project or HMI/SCADA data, and resulting operational disruption and financial loss. | Iranian-affiliated actor assessment. The advisory does not establish AI use. |
| Water utility controller attacks | July 2026; FBI/EPA warning July 30 | Utilities in at least seven states reported activity involving exposed Rockwell MicroLogix controllers. Attackers altered passwords and network addresses, causing loss of monitoring or control and some operational disruption. | Confirmed malicious activity; no verified AI attribution in the cited warning. |
| AI-generated scripts targeting PLCs | Canadian advisory updated September 14, 2026; cites U.S. partner warnings from August | Government agencies described AI-generated exploitation scripts and targeted reconnaissance against exposed industrial controllers, including Siemens S7 environments. | AI involvement in offensive tooling is supported. It does not establish an AI-caused national electricity outage. |
| UNC6780 / TeamPCP development supply chain | 2026 campaign; Google report September 8 | Google documented compromised developer accounts, trojanized MCP packages, altered repositories, and DUSTMAKER functionality targeting development and publishing workflows. | Google attributes the activity to UNC6780. Observed malware capabilities and individual confirmed outcomes must remain distinguishable. |
| MIDNIGHT NEPTUNE / SOMBERMEME | Discussed in Google’s November 16 investigation | Suspected DPRK-linked operations abused Claude Code repository settings as an execution path to deliver a backdoor after initial compromise. | Distinct actor tracking and attack path; no established identity with UNC6780 or testmetest operators. |
| testmetest / GitHub repository compromise | JFrog report October 26 | Researchers found 42 armed repositories across 11 legitimate owners, malicious development hooks, and connected package-publishing infrastructure. | Confirmed repository compromise and campaign tooling, not 42 proven successful developer-device infections. |
| NSA network degradation | November 17 | NSA reported temporary degradation during scheduled routine maintenance and said mission-critical capabilities remained operational. The agency statement does not name the affected software, a CVE, or a malicious cause. | Not included as a confirmed cyberattack. No public evidence establishes malicious exploitation as the cause. |
The last record belongs in an enterprise security evidence tracker, but not in the same incident count as malicious intrusions.
That distinction is intentional.
A vulnerability can create a security emergency. The urgent change intended to fix it can cause operational disruption. Both outcomes may require incident management, but only one involves malicious exploitation.
Likewise, the Canadian warning referenced earlier industrial disruption while describing AI-generated exploitation tooling. That does not retroactively establish that every previous industrial incident was AI-driven.
This is why the tracker needs separate fields for attack activity, AI involvement, attribution, and operational effect.
Which Incidents Are Actually Connected?
A useful tracker should support relationships between records without assuming they share an operator.
The strength of a connection depends on the evidence.
Confirmed or Directly Documented Connections
Storm-0940, ARTEX, and ConnectWise Cloud Backup: Microsoft directly connects the tracked actor, use of ARTEX, compromised application identities, attempted tenant authentication, and some follow-on Microsoft Graph activity. These belong in one documented campaign relationship. See Microsoft’s investigation.
RMM+, ScreenConnect, and SCKit: ConnectWise and Huntress connect the third-party extension exposure, suspicious management commands, and unauthorized client deployment. The separately vulnerable technician-client execution path is relevant to the toolkit’s intended propagation, but the complete SCKit propagation sequence was not observed succeeding. See ConnectWise’s RMM+ disclosure.
Optimizely account impersonation and malicious JavaScript delivery: Optimizely directly connects the signup-flow flaw to unauthorized Project JavaScript changes. Report URI independently confirms contaminated snippets and browser-side malicious behavior. Those findings support one incident record with multiple evidence sources. See Report URI’s independent observations.
testmetest packages and compromised GitHub repositories: JFrog establishes operational links through shared payloads, repository hooks, CI publishing components, and related infrastructure. Its investigation also connects elements of the repository cluster to previously reported shellpanel activity. It does not establish successful exploitation of every downstream target or a single uninterrupted infection chain. See JFrog’s repository analysis.
PHP2Shell vulnerabilities and observed exploitation: CISA’s KEV additions and vendor reports establish exploitation of the two PHP vulnerabilities. This is a supported vulnerability-to-exploitation relationship. It does not assign every exploitation attempt to one actor. See CISA’s KEV additions.
Shared Techniques, but No Established Common Operator
Several incidents have important architectural relationships without established actor identity.
| Records | Shared pattern | What is not established |
|---|---|---|
| ARTEX / Storm-0940 and CAIRN / CL-STA-1238 | Agentic attack orchestration and cloud access expansion | Same operator, shared command structure, or one campaign |
| Storm-0940 and BRICKSTORM | China-nexus reporting and abuse of privileged infrastructure | Same threat actor or coordinated intrusion |
| exploit_agent and PHP2Shell | AI-assisted vulnerability research and rapidly developed exploit capability | That the recovered exploit_agent operator conducted every PHP2Shell exploitation campaign |
| ConnectWise Cloud Backup and ScreenConnect | Same supplier and downstream customer trust risks | Same intrusion, affected component, or initial access mechanism |
| Optimizely and compromised MCP integrations | Trusted third-party execution paths | Same attacker, malware infrastructure, or root cause |
| AI-assisted PLC targeting and U.S. water utility disruptions | Exposed industrial controllers and unauthorized control-system access | That the documented operational disruptions were all AI-assisted |
| UNC6780, MIDNIGHT NEPTUNE, and testmetest | Developer tooling, repository configuration, and executable supply-chain abuse | A single actor or unified campaign |
This distinction should be preserved in diagrams and security reports.
A direct connection can justify a shared investigation scope.
A shared technique usually justifies a common defensive control.
It does not automatically justify treating several actor groups as one adversary or aggregating their victim counts.
A Connection Map That Does Not Invent a Common Adversary
The following diagram distinguishes source-supported campaign relationships from recurring technical patterns.

The solid arrows represent relationships documented by the investigators who tracked those campaigns. They do not establish that every attempted operation or analyzed malware capability succeeded.
The second group represents an architectural synthesis.
The synthesis is useful because different incidents can justify similar defensive improvements without belonging to one threat operation.
That is how a cross-incident tracker should inform enterprise architecture.
Why the Reported Numbers Cannot Be Added Together
A common analytical mistake is to sum numbers taken from unrelated research reports and describe the result as total affected organizations.
The denominator may be different in every source.
| Reported measure | Correct unit | What must not be inferred |
|---|---|---|
| More than 2,250 ConnectWise customer tenants targeted | Attempted tenant authentication | 2,250 confirmed tenant compromises |
| CAIRN’s 25 operations against at least 19 organizations | Operations and targeted organizations | 25 compromised organizations |
| CAIRN’s 10 confirmed successful targets | Organizations with unauthorized access or retrieval | 10 identical incident outcomes |
| BRICKSTORM’s 393-day average | Mean undetected dwell time within Mandiant’s investigated intrusions | A universal cyberattack dwell-time benchmark |
| SCKit’s 288 and 205 affected hosts | Hosts in two reported observation waves | A proven deduplicated total of unique organizations |
| At least 688,000 guest-originated ScreenConnect file-execution requests | Attempted file-execution requests | 688,000 successful executions |
| Optimizely’s 188 infected snippets and 178 organizations | Contaminated script artifacts and associated organizations | 178 breached corporate networks |
| testmetest’s 42 armed repositories | Compromised or prepared repository objects | 42 confirmed workstation infections |
| U.S. water utility incidents in at least seven states | Geographic scope of reported incidents | Every utility in those states was compromised |
The distinctions are not editorial technicalities.
They determine the appropriate scale of remediation, notifications, customer investigation, resource allocation, and executive reporting.
For example, every customer tenant with a previously authorized application may need an exposure review. That does not mean the organization should report confirmed data theft before the available evidence establishes it.
Likewise, a large number of blocked or unsuccessful exploitation attempts can justify urgent defensive action without being presented as an equivalent number of successful intrusions.
A useful tracker preserves the original unit, observation period, population, and source limitations next to each number.
In its November 18 update, Optimizely reported that fewer than 5% of its customers were impacted and that remediation was complete on its side. That vendor customer percentage uses a different population from Report URI’s snippet and organization counts. Customer-side investigation of the exposure window remains a separate task, and the update does not establish AI involvement in this incident.
What AI Involvement Can Actually Be Claimed?
One of the central questions running through this series is whether AI is changing cyber operations.
The evidence supports that conclusion, but AI involvement must be classified by its actual role.
| AI involvement | Cases supported by the reviewed evidence | Interpretation |
|---|---|---|
| Observed agentic intrusion orchestration | ARTEX in Storm-0940 activity; CAIRN in the CL-STA-1238 investigation | AI-driven frameworks performed multiple operational tasks during documented intrusion campaigns |
| Observed automated exploit-research infrastructure | exploit_agent | Researchers recovered attacker infrastructure configured for automated vulnerability analysis and exploitation |
| AI-assisted social engineering or development | MIDNIGHT NEPTUNE; Canadian industrial-controller reporting | AI-supported content generation or tooling is documented, with different roles in the attack |
| AI tooling abused as an execution surface | Claude Code configuration abuse; compromised MCP integrations; development hooks | The trusted development tool or integration can execute malicious configuration or software |
| Possible AI-generated malware components | SCKit RelayTrap code characteristics discussed by Huntress | Consistent with suspected AI-assisted development, not proof of autonomous AI operation |
| AI involvement not established | Optimizely injection, the July water utility intrusions, BRICKSTORM, and the NSA maintenance event | The reviewed source evidence does not justify an AI-driven attribution for those events |
This classification prevents two opposite errors.
The first is dismissing meaningful AI-enabled operations simply because the final attacker action used ordinary credentials, scripts, or APIs.
The second is calling every automated attack AI-driven.
A Python script that sends thousands of requests is automation.
An LLM that generates an exploitation script is AI-assisted development.
A framework that uses models, tools, state, and execution feedback to perform multiple intrusion stages is an agentic workflow.
Those differences affect detection, investigation, and the controls that can constrain the activity.
They should remain visible in the tracker.
The Architectural Connection: Attackers Keep Crossing Established Trust Boundaries
The most consequential relationship among these cases is that attackers repeatedly used mechanisms organizations already trusted.
The vulnerabilities and affected products differed, but several patterns recur.
Nonhuman Identity Has Become a Critical Privilege Boundary
The ConnectWise Cloud Backup compromise demonstrates the reach of supplier-managed application credentials.
CAIRN demonstrates the creation of persistent administrative accounts and API authority.
Developer supply-chain investigations show the consequences of compromising publishing identities and automation credentials.
These cases support a common enterprise requirement: maintain an inventory of nonhuman identities and the actual permissions they exercise.
The inventory should distinguish the credential issuer, owning application, target resources, administrative privileges, and independent restriction path.
Rotating a compromised credential may remove the original access mechanism. It does not prove that another credential or service identity created during the intrusion has been removed.
Management Software Can Become an Access Multiplier
BRICKSTORM targeted infrastructure management appliances.
SCKit exploited remote-management authority to deploy unauthorized clients.
The industrial-controller cases targeted equipment and interfaces that govern physical processes.
Those environments differ operationally, but management access has similar security characteristics.
It can reach systems that are not normally exposed to the attacker.
It may authorize commands without each endpoint or operator granting a new approval.
And its compromise can affect the tools used to investigate and recover.
The defensive implication is to separate ordinary service operation from high-impact management authority, limit the reach of privileged identities, and preserve an independent way to investigate or restrict affected systems.
Trusted Software Can Carry Untrusted Execution
Optimizely’s Project JavaScript was delivered through legitimate script infrastructure.
The development supply-chain campaigns used recognized packages, repository files, and supported execution hooks.
SCKit installed unauthorized ScreenConnect clients that could resemble familiar administration software.
A recognized product name, valid certificate, approved package source, or trusted repository does not prove that the current executable content was authorized.
The stronger control is to connect software provenance and change history to execution permissions.
For browser-side dependencies, that includes script monitoring and carefully designed Content Security Policy enforcement.
For developer tooling, it includes protected configuration, package provenance, MCP server authorization, and restricted execution environments.
For infrastructure management, it includes verified software sources, approved services, and independently retained configuration evidence.
Closing the Entry Point Is Not the Same as Eradication
This distinction appears repeatedly throughout the series.
Patching a PHP vulnerability does not remove a webshell or compromised credential that an attacker established earlier.
Disabling RMM+ does not automatically remove unauthorized ScreenConnect clients already installed on endpoints.
Revoking supplier application credentials does not reverse unauthorized permissions or accounts created in customer tenants.
Rebuilding a management appliance does not necessarily remove altered automation or a second privileged identity.
The response must establish which downstream authority remains and how it can be verified independently.
This is why persistence and recovery belong in the same architecture discussion.
Turn the Tracker Into a Decision System
A threat tracker is useful only when it changes what operators investigate, restrict, or validate.
A collection of incident headlines may improve awareness without improving readiness.
A better approach is to map each evidence class to an accountable enterprise decision.
| Enterprise exposure | Decision supported by this tracker | Required evidence |
|---|---|---|
| Third-party cloud applications | Which supplier identities can access critical resources, and can they be restricted independently? | Service-principal inventory, permissions, authentication and downstream audit records |
| Internet-facing vulnerable software | Which affected runtime or appliance has confirmed exposure and a supported fix? | Active component inventory, vendor advisory, patch verification and compromise assessment |
| Remote-management tools | Can unauthorized commands, extensions, or clients persist after containment? | Management logs, extension state, client provenance, approved relay inventory |
| Browser-executed third-party code | Can suppliers change executable behavior without an application deployment? | Script inventory, version history, browser policy and historical connection evidence |
| Industrial control systems | Which external paths can reach controllers, and what physical consequences are possible? | OT asset records, approved remote access, engineering configuration and process dependencies |
| AI development environments | Can repository content, MCP packages, or hooks reach privileged local or CI/CD execution? | Protected configuration, tool authorization, runner permissions and publishing provenance |
| Infrastructure and recovery | Can compromised administration alter both production and recovery state? | Privileged identity graph, independent logs, protected copies and restoration tests |
| Security-driven maintenance | Can an emergency patch be applied without losing control of critical service dependencies? | Supported change plan, stop conditions, fallback and end-to-end acceptance evidence |
These decisions can be assigned to existing service, identity, security, application, platform, and recovery owners.
They do not require one security platform to become the authoritative source for every environment.
They require a reliable relationship between the incident intelligence and the enterprise assets that might be exposed.
A Practical Evidence Record
For ongoing maintenance, each tracker entry should use a consistent structure.
The following YAML illustrates a conceptual tracking record for the ConnectWise Cloud Backup investigation. It is not a feed from Microsoft, a threat-intelligence platform, or an automatically maintained system.
record:
id: DTD-2026-STORM0940-CONNECTWISE
snapshot_date: "2026-11-18"
classification:
type: confirmed_intrusion_activity
ai_role: observed_agentic_orchestration
campaign:
source_actor_name: Storm-0940
attribution_source: Microsoft
attribution_description: China-state-aligned
framework: ARTEX
timeline:
activity_period: "2026-09 through 2026-10"
source_publication: "2026-10-21"
evidence:
observed:
- attempted_authentication_to_customer_tenants
- follow_on_graph_activity_in_some_environments
- credential_and_privilege_modifications
reported_reach:
measure: attempted_customer_tenants
count: ">2250"
unproven:
- all_targeted_tenants_compromised
- all_targeted_tenants_exfiltrated_data
relationships:
confirmed:
- compromised_connectwise_application_trust
- observed_use_of_artex
not_established:
- same_actor_as_cl_sta_1238
- same_intrusion_as_screenconnect_sckit
enterprise_response:
scope: affected_application_identities_and_tenants
validate:
- service_principal_sign_ins
- graph_and_resource_activity
- unauthorized_credentials_and_grants
- downstream_access_after_revocation
maintenance:
review_trigger:
- new_vendor_disclosure
- new_attribution_evidence
- changed_impact_assessment
- confirmed_customer_remediation
The structure separates observed facts from unsupported claims.
It also distinguishes the reporting organization’s actor designation from the organization maintaining the tracker.
The purpose is not to create a permanent, unquestionable attribution.
It is to provide a current evidence-backed assessment that can be revised without losing its provenance.
Preserve the Source History
When a vendor updates an advisory, the tracker should retain the previous source date and relevant earlier findings.
A later report may narrow a victim population, identify a new execution path, correct a technical assumption, or change attribution.
Do not silently overwrite the earlier position.
Record the new finding, the evidence that changed it, and the operational decision affected by the change.
A source correction may justify reopening a previously closed investigation.
Similarly, a new remediation release may change what platform owners are required to deploy without changing the historical facts of an intrusion.
Make Unknowns Explicit
Every incident record should distinguish:
Unknown: No sufficient evidence is available to support the claim.
Not observed: The relevant evidence source was examined, and the specific activity was not identified within its scope and period.
Not applicable: The mechanism does not apply to the affected configuration or component.
Confirmed absent within scope: An appropriate validation established that the specific condition was not present in the defined environment.
These labels are not interchangeable.
For example, a tenant whose October service-principal logs have expired cannot claim no suspicious October authentication occurred merely because no records appear in the current log view.
Likewise, an organization that was never running a vulnerable third-party RMM+ extension should not be classified as remediated by a ScreenConnect server update.
The state of the evidence should match the decision being made.
How to Keep This Tracker Current
This tracker provides a November 18, 2026 evidence baseline, and the incidents will continue to develop.
The tracker should be maintained through explicit review triggers rather than indefinite assumptions about current status.
A new vendor advisory may identify additional affected releases.
An incident-response firm may report previously unknown successful exploitation.
An attribution assessment may change when new evidence links infrastructure, tooling, or command relationships.
A supplier may complete containment while individual customer environments remain under investigation.
A government agency may publish a final root-cause report that resolves earlier operational uncertainty.
Each event should trigger a review of the relevant record, its connections, and the control actions derived from it.
The practical maintenance cycle is:

A useful update should answer more than what changed in the news.
It should explain whether the new information changes exposure, confidence, remediation, or recovery requirements for the organization.
That is the difference between a threat-intelligence feed and an operational decision system.
Conclusion
The evidence reviewed across this series shows a real evolution in cyber operations.
Agentic frameworks are being used in documented intrusions. Attackers are experimenting with automated vulnerability research and exploitation. Trusted cloud applications, management platforms, developer tools, and browser dependencies continue to provide consequential attack paths. Industrial controllers remain vulnerable when their remote access is poorly governed.
Those developments justify significant defensive attention.
They do not justify merging every campaign into one actor, describing every automated action as AI-driven, converting attempted authentication into confirmed breaches, or treating a maintenance outage as a malicious cyberattack without evidence.
Across this series, the durable architectural lesson is consistent.
The enterprise must understand the authority it has granted, the boundaries that authority can cross, the evidence generated when it is exercised, and the independent controls needed to remove it.
The next practical action is to select one critical business service and map its privileged identities, supplier integrations, remote-management connections, executable dependencies, and recovery authority.
Then use the confirmed incidents in this tracker to test whether those boundaries would withstand the corresponding failure modes.
An effective threat tracker does not simply tell an organization which attackers are active.
It helps the organization determine whether those attackers’ observed methods can succeed in its own environment, which controls matter, and what evidence would prove the controls are working.
External References
- Microsoft Security Blog: AI-driven, large-scale nation-state campaign using ARTEX enables supply chain compromise
- Google Threat Intelligence Group: ARTEX and Friends: Enter Agentic Offensive Frameworks
- Unit 42: Attackers Use Autonomous AI to Target Semiconductor, Defense and Civil Society Sectors
- Google Threat Intelligence Group: GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold
- Google Threat Intelligence Group and Mandiant: Another BRICKSTORM: Stealthy Backdoor Enabling Espionage into Tech and Legal Sectors
- CISA: CISA Adds Two Known Exploited Vulnerabilities to Catalog
- Palo Alto Networks: CVE-2026-109618 Impact of Critical Vulnerabilities in PHP on Palo Alto Networks Products
- ConnectWise: Trust Center | Advisories
- ConnectWise: CVE-2026-127063, Third-Party RMM+ Extension
- ConnectWise: CVE-2026-84869, ScreenConnect Client
- Huntress: SCKit Worm Installs Rogue ScreenConnect Clients and Session Cookie Relays Across Hundreds of Endpoints
- Huntress: CVE-2026-84869 in ScreenConnect Exploited to Run Malicious Files on MSP Technician Accounts
- Optimizely: Security Incident Initial Report for Customers, Web Experimentation, November 17, 2026
- Report URI: World’s Largest Brands Infected With Malware Via Optimizely
- Canadian Centre for Cyber Security: Cyber threat actors use artificial intelligence in an active global campaign to disrupt internet-exposed programmable logic controllers
- CISA: Iranian-Affiliated Cyber Actors Exploit Programmable Logic Controllers Across US Critical Infrastructure
- FBI: Malicious Cyber Actors Targeting Water and Wastewater Sector Internet-Facing Programmable Logic Controllers, Causing Operational Disruptions
- Google Threat Intelligence Group: From Prompting to Autonomy: The Evolution of Adversarial AI
- JFrog Security Research: Beyond testmetest: Native Addon Camouflage and 42 Armed GitHub Repositories Reveal a Hidden npm Publisher
- National Security Agency: NSA statement on network issues
- FBI, CISA, NSA, EPA, DOE, and CNMF: Joint PLC advisory AA26-097A, April 7, 2026 (FBI-hosted PDF)
- Optimizely: Web Experimentation Security Incident, November 18 update