The new era of cybersecurity: when the scale of attacks outpaces the capacity to defend
A flaw becomes an attack tool in under 24 hours; fixing it can take 30 to 60 days. AI agents and compromised vendors are widening that gap.
The digital infrastructure that underpins the global economy is entering a new phase of vulnerability. Companies, governments, financial institutions, hospitals, industrial systems and essential services depend on increasingly interconnected networks, while the ability to exploit their weaknesses is growing faster than many organizations can identify, contain and fix them.
This is not just a rise in the number of attacks. What is changing is the relationship between offensive and defensive capability. Vulnerabilities can be discovered, exploited and folded into criminal operations in ever shorter intervals. At the same time, the complexity of corporate infrastructure, the reliance on outside vendors and the proliferation of connected services widen the paths along which a compromise can spread.
The data help gauge the scale of this shift. The Microsoft Digital Defense Report 2026, published in October, counted roughly 40,000 vulnerabilities (CVEs) disclosed in the first half of the year. The median time between the discovery of a vulnerability under exploitation and its conversion into an attack tool fell to under 24 hours. Fixing critical, externally exposed vulnerabilities, by contrast, can take 30 to 60 days in corporate environments.
The asymmetry is plain. An attacker can work around the clock, test alternatives, automate procedures and reuse tools against different targets. Defense, in many organizations, relies on fragmented processes, limited staff, update cycles and response structures that do not move at the same speed.
That gap takes on another dimension with technological interdependence. According to Verizon's Data Breach Investigations Report 2026, third parties were involved in 48% of the breaches analyzed, up from 30% in the previous edition. The figure covers different forms of involvement, including software vulnerabilities and vendor compromises. An intrusion does not have to begin at the company that suffers the consequences: it can come in through a software vendor, an authorized application, a compromised credential or a service that holds access to hundreds or thousands of environments.
Artificial intelligence adds a decisive variable. It speeds up reconnaissance, vulnerability analysis, tool development, social engineering and the processing of information gathered during an intrusion. More recently, frameworks built on specialized agents and coordination mechanisms have begun to carry out long operational sequences with less continuous human involvement. This does not mean every attack is autonomous or AI-driven. It means automation is shifting the limits of productivity and scale available to offensive actors.
The central argument has three parts. First, AI can lower the cost of, and the need for human involvement in, offensive operations. Second, what turns models into operational capability is the harness, the software layer that connects a model to tools, data, memory and permissions. Third, the supply chain extends the reach of any compromise. The cases below are meant to demonstrate each of these claims, with the caveats the sources require.
This essay examines four developments documented in October and November 2026: offensive automation in real operations, with ARTEX as the central case; the architecture that supports it; the chain of trust between vendors, which provides the ground where this capability meets existing vulnerabilities; and the geopolitical backdrop, in which states, espionage groups and criminal organizations exploit the same dependencies.
The consequences go beyond information technology. When a software company is compromised, its customers can become indirect victims. When a cloud service is hit, business and administrative operations can stop. When hospitals, energy systems, transport or communications networks are struck, the effects are no longer purely digital and reach everyday life. The question is no longer only how to stop an intrusion. It is how to protect a society whose economy, public administration and communications depend on an interconnected infrastructure that is permanently exposed and ever more complex.
ARTEX: from assistive AI to operational AI
For a long time, artificial intelligence played a supporting role in cyber activity. Language models interpreted code, researched vulnerabilities, wrote scripts and translated technical documentation. A human still organized the tasks, ran the tools, interpreted the results and decided the next move.
That relationship is changing. A new generation of platforms combines language models, specialized agents, external tools and coordination mechanisms into frameworks able to carry out entire sequences of activity. Instead of answering a request, they receive a goal, break the work into smaller tasks, invoke tools, analyze the results and decide on the next action. A model can explain how to find a vulnerability. An agent system can test an environment, run analysis tools, interpret the results and move on. Used maliciously, the same capability is turned against other people's systems.
ARTEX is the most concrete case of this shift. Written in Go with a web interface, it was designed as a security testing platform so that technical teams could assess systems under their responsibility. The change came when malicious actors began using it against organizations they were not authorized to test.
On November 13, 2026, the FBI issued alert FLASH-20261113-001 about the tool. According to the document, various malicious actors have used it since at least August 2026 to automate stages of intrusions, from initial reconnaissance to data exfiltration. The campaigns hit financial services, government, technology, defense, semiconductors, energy, health care, manufacturing and telecommunications. Among the highlighted cases are attacks on South Korean financial institutions, with personal data of about 67,000 people exfiltrated, and the supply-chain compromise of ConnectWise Cloud Backup, which reportedly enabled access to more than 2,250 downstream customer environments. The most relevant figure is speed: according to the FBI, some intrusions went from initial targeting to privileged access in under three hours.
The platform has no language model of its own. It connects external models, commercial and open-weight, to a framework of specialized agents, takes goals described in natural language and turns them into executable tasks using security tools and coordination mechanisms. This allows one model to be swapped for another without rebuilding the framework. The model supplies interpretation and reasoning; the architecture organizes the work, manages the tools and tracks execution. That is why ARTEX's operational capability is not tied to any single AI provider.
The original developer took the public repository off GitHub on October 8, after the malicious use gained visibility. Derived versions, copies and community forks remained available. Removing the original code did not remove the tool from the offensive ecosystem.
Two weeks and 36 billion tokens
On November 5, Check Point Research published Hacking on Autopilot: Inside ARTEX Cybercrime Campaigns. What set the study apart was access to internal records from ten misconfigured installations, whose databases exposed usage logs, configured models and interactions between operators and agents. The researchers identified 2,557 systems that hosted or had hosted ARTEX panels, of which 1,608 were active and publicly reachable at the time of the survey. The number does not represent all installations, since many sit on private networks or behind firewalls. Among the instances examined, Chinese models predominated, especially those of the DeepSeek Flash family.
One of the documented cases is revealing. Between October 14 and 29, 2026, an operator ran two panels at the same time, with the DeepSeek V4.1 Flash and Kimi K3.1 Pro models. The goal was to obtain digital certificates and passwords from ten companies that sell app-signing certificates for Apple devices. These certificates allow certain software to be distributed outside the usual App Store channels, and obtaining them illicitly can make it easier to distribute tampered apps.
Over the period, the systems processed about 36 billion tokens:
- 2.13 billion uncached input tokens;
- 33.79 billion cached input tokens;
- 473.9 million output tokens.
Based on DeepSeek's official API prices at the time, Check Point estimated the cost of model usage at between $700 and $1,500. According to Check Point, the agents obtained more than 300 certificates along with their passwords. After validation, three met the technical criteria for use in app signing.
Human involvement was small. The logs show an operator who asked about task progress, requested information on targets, checked which ones were closest to the goal and adjusted priorities. There was no need to follow each attempt. Human supervision still existed, but it was no longer needed at every step: the operator set goals and tracked progress, while the agents did much of the intermediate work.
The counterpoint: agents that get it wrong
The same investigation includes a case with a different outcome. In another campaign, an operator used ARTEX to try to obtain cryptocurrency from platforms and financial services. The agents were assigned tasks across 292 payment-infrastructure domains and processed about 167.5 million tokens in four days, using DeepSeek V4 Pro, at an estimated cost of $145 to $300. They found vulnerabilities, accessed information and produced reports on exploitation opportunities, but did not achieve the financial goal. At one point, they read API responses as confirmation of successful transactions. Later checks showed no movement in the operator's wallets.
The episode exposes a limitation of agent-based systems. Carrying out tasks, interpreting responses and producing reports does not mean correctly understanding the result of one's own actions. An agent can identify a real vulnerability and at the same time misjudge whether an exploit succeeded. Automation also multiplies errors. Speed does not remove the need for verification.
The machine is the architecture
Discussion of AI and security tends to focus on models: which is smartest, which writes the best code, which has the fewest restrictions. A decisive piece gets left out: the software layer that turns a model into an operator. That layer is the harness. In September 2026, the Australian Cyber Security Centre defined it as the layer that connects language models to tools, data sources, memory and planning workflows, while also controlling privileges, managing execution and recording metrics.
A model on its own takes in information and produces answers. It has no automatic access to external systems, no persistent operational memory and no ability to run a task indefinitely. The model determines part of the reasoning capability; the harness determines how it is used, what resources it can reach and how actions are chained together. Six components define how efficient this architecture is:
- Orchestration. Turns a broad goal into smaller tasks, distributes them among agents and tracks the results, either hierarchically or collaboratively. Google Threat Intelligence Group (GTIG) described frameworks such as CAIRN, CyberStrikeAI and PentAGI, some built around observe, orient, decide and act loops.
- Model routing. Picks the model for each task based on cost per token, speed, reasoning capability, availability and usage limits. It allows an unavailable model to be replaced without halting the operation.
- Tool execution. Converts model decisions into actions. GTIG describes frameworks that connect models to dozens or hundreds of security tools.
- Memory and persistence. Keeps earlier results, pending tasks, decisions and errors, so the operation can resume without rebuilding context.
- Supervision and validation. Assesses whether a task was completed and whether the result is reliable, through a supervisor agent, a verifier, deterministic rules or human intervention.
- Resource management. Handles API calls, token consumption, caching, parallel execution, storage, infrastructure and provider limits.
Routing and economics
Using different models within one operation opens up an economic possibility: there is no need to use the most expensive model for every task. Google Cloud's architecture documentation describes the strategy of choosing models by complexity, latency and inference cost, and notes that multi-agent systems spread responsibilities but add processing, communication and coordination costs.
A hypothetical architecture would have a more capable supervisor model to plan and evaluate; cheaper executor models for high-volume tasks; verifier models to catch inconsistencies; and an orchestrator to distribute the work. One can imagine Kimi K3.1 Pro in the supervisor role and DeepSeek V4.1 Flash doing the execution.
The $700 to $1,500 figure is an estimate of inference cost, not the budget of a campaign. It leaves out servers, hosting, proxies, purchased access, harness development and upkeep, outside services and monitoring. Token consumption also does not grow linearly: agents repeat steps, exchange messages and reload long contexts. In the case analyzed, almost 34 billion of the roughly 36 billion tokens were cached input, billed differently from raw volume. Even so, the figure suggests AI processing may be a relatively small share of an operation's cost. It does not mean any sophisticated operation can be run for that amount. Automation does not make every attack cheap, but it changes the ratio between investment and offensive capability: it lowers the marginal cost of new attempts and makes it possible to hit more targets and sustain operations for longer.
The new risk surface
The layer that extends operational capability also concentrates risk. A harness can have access to credentials, files, corporate systems and tools with different privilege levels. If those permissions are excessive or poorly managed, compromising the agent has consequences that go beyond the model. The Australian guidance stresses that the significant risks arise from the connections between models, tools, data and external systems, and recommends least privilege, identity management, monitoring and human oversight. A system can use a model with sophisticated safeguards and still be vulnerable in how it handles tools, credentials and permissions.
The advantage, then, does not necessarily lie in owning the most advanced model. It lies in building the architecture that gets the most out of each model, at the right moment, at the lowest operating cost.
An ecosystem, not a tool
ARTEX is not an isolated phenomenon. On October 30, GTIG, working with Mandiant, published an investigation into agent-based offensive frameworks, identifying the use of ARTEX, HexStrike and CAIRN in real-world activity. According to the researchers, these frameworks were already automating stages of the intrusion cycle, such as lateral movement, database exploitation and preparation for data exfiltration. On November 16, an update broadened the survey, covering the use of CAIRN and PentAGI in real operations and experimentation with CyberStrikeAI and HexStrike AI.
- CAIRN. A multi-agent framework built on the OODA loop (observe, orient, decide, act). Agents share information in a common environment. Google identified its use in intrusions against targets that include U.S. aerospace organizations.
- PentAGI. An offensive automation platform designed to lower the technical barriers to complex campaigns. Between late August and early October, Google tracked a threat actor that used a customized version to carry out systematic reconnaissance and vulnerability scanning against organizations in various sectors across Asia.
- CyberStrikeAI. Connects commercial and open-weight models to more than a hundred security tools, with a coordination layer that organizes the stages of an operation.
- HexStrike AI. Connects language models to more than 150 offensive security tools, covering reconnaissance, web application analysis, exploitation and binary analysis, with more than a dozen specialized subagents.
Google also identified custom harnesses built by threat groups and pointed to in-house frameworks developed by groups linked to cyber espionage.
The key point is diversity. There is no single agent model and no dominant platform. Models, agents, tools, inference services and coordination frameworks can be combined in many ways. The model can change, the tool can be replaced, new agents can be added. The architecture remains the organizing element, and offensive capability can be assembled from relatively accessible components.
From patch to exploit
Google also reported the use of automated systems to compare software versions, especially when an update fixes a vulnerability. This technique, patch diffing, examines what changed in the code to work out which flaw was fixed. From there, agents can try to develop proofs of concept and prepare exploits against systems that have not yet installed the fix. In November, Google described an automated research environment that combined continuous patch diffing, vulnerability identification, exploit development and test execution without human involvement at certain stages.
The potential consequence is a shorter interval between the release of a fix and the exploitation of those who have not applied it. The report describes agents that explore multiple attack paths, fix execution errors and adjust strategy in seconds or minutes. But speed has a cost: without constant supervision, operations generate heavy traffic, recognizable identifiers and technical traces tied to the tools, which can make detection easier. Offensive speed does not mean invisibility.
The other side: AI that finds vulnerabilities
AI's offensive capability cannot be understood without the same technology's potential to discover and fix flaws in critical systems. On November 4, 2026, OpenAI published the results of research on using cybersecurity-specialized models to find vulnerabilities in the Linux kernel. The experiment set up an automated pipeline combining code analysis, dynamic reproduction of vulnerabilities, exploitability assessment and generation of candidate fixes.
In a single run, in a controlled environment, the system identified and analyzed 2,308 potential issues, dynamically reproduced 1,724 vulnerabilities and produced 948 working local privilege escalation exploits, a type of flaw that can let a user with limited access gain administrative privileges, up to root (that is, administrator level). The reported cost was about $1.21 million in API credits, or roughly $1,279 per exploit. OpenAI also noted that more capable models find flaws that earlier versions missed, even with the same analysis process.
The numbers call for careful reading. The research does not show that 948 Linux servers were breached, nor that 948 vulnerabilities were exploited in production. These are results obtained on controlled infrastructure, with later validation and review by human researchers. During the ten days of the run, more than a hundred local privilege escalation vulnerabilities had already been fixed, or had patches being prepared by other contributors to the Linux project.
Even so, the demonstration matters. Research that once required specialized teams working for long periods can be carried out by an automated, systematic process, and the ability to find vulnerabilities itself becomes more scalable. The consequence is unavoidable: techniques that can identify flaws, reproduce them and develop proofs of concept can also be used by malicious actors. Technology does not determine the purpose of an operation. Context, permissions, safeguards and control over the tools are what decide it. And there is a significant difference between finding a flaw, exploiting it and achieving an actual compromise. Together, ARTEX and the OpenAI research support the thesis that AI is changing productivity on both sides of the contest.
The chain of trust: the weakest link outside the organization
Traditional corporate security starts from the idea of protecting one's own systems. That model hits its limits when a growing share of infrastructure depends on third-party services: storage, backup, communications, authentication, management and remote administration. To work, these services receive permissions and credentials that give them access to customers' internal environments. The vendor already has legitimate access, and if those permissions are compromised, the intruder inherits part of the trust placed in the provider, even at organizations with robust security.
ConnectWise: the vendor as an access vector
On October 14, 2026, ConnectWise opened an investigation after detecting unauthorized activity tied to the application credentials of Cloud Backup, formerly SkyKick, and suspended services to contain the problem. The service is integrated into partners' and customers' Microsoft 365 environments and relies on application identities and permissions granted by administrators. On October 17, the company said it had disabled seven application identities and revoked certificates, and advised partners to review permissions, suspicious sign-ins and unrecognized access. The suspension halted scheduled backups, file access and data restoration, and also affected SaaS Security and Cloud Management. Backup files, according to the company, remained encrypted and isolated, with no evidence of compromise.
Microsoft attributed the campaign to the group Storm-0940, which it links to Chinese state interests, and to the use of ARTEX to automate reconnaissance, authentication attempts and exploitation of the resulting access. It reported attempts to reach more than 2,250 customer environments; the FBI speaks of access to more than 2,250 downstream customer environments. The numbers need care. In a password-spraying campaign between October 10 and 14, Microsoft logged more than 151,000 failed attempts against 327 accounts at eight organizations, with no successful sign-ins at that stage. In later activity tied to the compromised applications, it identified sign-ins and actions in some customer environments, including Microsoft Graph operations and changes that could enable persistence. Gaining access to an application identity, reaching an environment and extracting information are different steps. What the case shows is that an intruder does not need to create a new vulnerability at each organization: it can exploit a trust relationship that already exists.
OPSSHIELD: the update that delivers the threat
On November 2, the update distribution infrastructure for cPGuard, OPSSHIELD's security product for Linux servers, was compromised. Using improperly obtained credentials, the intruder pushed a malicious script through the legitimate update channel. The company said the attacker accessed part of its billing and licensing infrastructure, obtained the publishing keys and distributed a malicious version for about 70 minutes. The affected versions run from 5.88.12 to 5.88.19. The script installed a backdoor and harvested credentials and files from the servers it reached. OPSSHIELD published guidance to update, investigate systems and rotate potentially exposed credentials and SSH keys.
On November 3, the company estimated that the intruder had retrieved 32 files from customer servers. This is a preliminary estimate of files retrieved, not the total number of servers that received or ran the update, and should not be read as a victim count. The case hits a sensitive layer: administrators install security tools to protect themselves, and those tools' update mechanisms are highly trusted and often hold administrative privileges.
ASOS: social engineering against third-party platforms
Not every vendor compromise depends on a sophisticated technical flaw. On October 6, British retailer ASOS confirmed unauthorized activity on third-party platforms used to communicate with customers, after shoppers received an unauthorized notification through the company's channels. The investigation found that the intruder posed as a trusted contact to obtain an employee's credentials. Customer names and contact details may have been accessed; the company found no compromise of payment data or passwords. ASOS said it had completed the investigation and taken additional measures, and the episode temporarily disrupted marketing and customer-relations activity. One exploited trust relationship was enough, without breaching the retailer's core infrastructure.
Miljödata: one vendor's failure reaches millions
In August 2025, Swedish IT provider Miljödata suffered an attack that led to unauthorized access to, and publication of, large volumes of personal data. According to the company, about 2.2 million people were affected, with personal identity numbers, contact details and sensitive records on sick leave, rehabilitation and incidents involving students. Its customers included most Swedish municipalities, several administrative regions, government agencies and private companies. On October 1, 2026, the Swedish data protection authority, IMY, announced a fine of 1.8 million Swedish kronor for insufficient technical and organizational measures, including inadequate checks when installing new software and no automated real-time monitoring. A specialized company can hold data on a population far larger than its size suggests.
Qubole and PwC: forgotten permissions and data in third-party hands
The Qubole case, disclosed on November 9, shows how inherited trust works. According to the published research, attackers used identity and access management credentials tied to the platform's AWS account to assume access roles that customers had set up in their own environments (cross-account IAM roles), a legitimate integration feature. Some of these permissions were still active even after customers had stopped using the platform. The attackers were able to enumerate resources, access environment variables, query secrets and pull data from S3 buckets. The total number of affected organizations was not disclosed. Taking a service out of use does not mean revoking the permissions it was granted, and a forgotten integration can remain a way in.
On October 23, PwC confirmed an incident at an outside vendor that hosts employee training records, after an extortion group claimed to have stolen 590 GB of data. The public information at the time did not allow the extent of the access or the claimed volume to be confirmed. An organization may not have been breached directly and still face the consequences of a compromise at a provider that holds information on its behalf. The trust relationship is technical, but also contractual, operational and legal.
Optimizely: third-party code that reaches the browser
The November 16 incident illustrates the risk of third-party code well. The Optimizely Web Experimentation platform, used to test and personalize websites, serves JavaScript loaded directly by visitors' browsers. A component controlled by an outside vendor thus runs code inside customers' pages. According to Report URI, an attacker managed to inject malicious JavaScript into customer projects: 188 compromised files, associated with 178 organizations, including the BBC, CNN, Zoom, Salesforce, HSBC, IKEA and Equifax, delivered through the platform's legitimate infrastructure.
The malicious code served several purposes: swapping cryptocurrency wallet addresses, capturing card data typed into forms, stealing cookies and taking screenshots. There were also mechanisms to steer iPhone and iPad users toward iOS exploitation tools, which does not prove the devices were compromised. The distinction is essential: the 188 files came from a sample of 1,611 scripts examined. That does not mean 178 organizations had their servers breached, nor that every visitor was harmed, and Reuters reported that it was not possible to determine how many people were affected. What happened was different: legitimate sites began distributing malicious code because they trusted an external component. The file arrived from a legitimate domain, with a valid certificate and apparently normal behavior.
Six forms of exposure
The cases are not equivalent, but they share one trait: the attack surface extends beyond the organization's administrative and technical boundaries.
- Privileged cross-organization access (ConnectWise): a vendor's application credentials open its customers' environments.
- Persistent integrations (Qubole): old permissions stay valid after the service ends.
- Code run by third parties (Optimizely): external scripts operate in users' browsers without customers' servers being compromised.
- Update channel (OPSSHIELD): the installed protection becomes a malware distribution channel.
- Social engineering against external platforms (ASOS): one human credential opens third-party services.
- Data concentration (Miljödata, PwC): one provider holds information on many organizations and people.
Concentration widens the reach of an incident. That does not mean every customer will be compromised when the vendor is breached: it depends on the permissions granted, the controls in place and the nature of the affected component. There is also an asymmetry of visibility. A company can closely monitor its own servers and have little ability to track changes in an external platform's code or to investigate a provider's infrastructure in real time.
AI speeds up the problem but does not explain every incident
Recent incidents do not show that all supply-chain attacks use AI. The link is documented in the ConnectWise case, by Microsoft and the FBI. In the public information consulted on Qubole and Optimizely, there is no equivalent evidence, and the same goes for the other cases. AI did not create the structural fragility of the supply chain: excessive trust relationships, persistent credentials, broad permissions and dependence on external components were already known problems. It can increase the speed and scale at which these weaknesses are found and exploited, but it should not become the automatic explanation for every incident. What is taking shape is a convergence of risks: the attack surface is growing while the cost of exploiting certain weaknesses may be falling.
Defense at machine speed
The idea that offensive speed outpaces defensive speed lacks precision. There is no single measure comparing the time of all attacks with that of all fixes: incidents vary in complexity, environments differ in maturity and not every vulnerability can be exploited under real conditions. The documented cases, however, show a worrying asymmetry. An agent can run reconnaissance in the middle of the night, test multiple access paths and keep working through holidays and weekends. An organization juggles legacy systems, vendor dependencies, maintenance windows and internal approval processes.
Fixing a vulnerability does not end when the vendor ships the update package. The exposed systems must be identified, compatibility assessed, side effects tested, the update rolled out and the fix verified. In critical infrastructure such as energy, transport, telecommunications and financial services, a poorly planned update can cause disruptions as serious as the ones it is meant to prevent. Speed of discovery does not automatically translate into speed of remediation.
The problem is also one of coordination. Facing an incident, an organization needs to know which systems were hit, how far the compromise extends, whether there is persistent access, what the operational risks are and which measures can be taken without disrupting essential services. A poor response in complex environments can create new consequences. Reliance on a few vendors for software, cloud, identity and authentication widens the potential reach of every incident, and the discussion comes to involve the dependency architecture of the entire digital economy.
What automated offense still cannot solve
Defense is not doomed to lose the race. Offensive frameworks have limits: automation does not remove the need for infrastructure, valid credentials, exploitable vulnerabilities and achievable goals, and agents make mistakes, as the cryptocurrency campaign showed. The heavy traffic, recognizable identifiers and technical traces Google pointed to create behavioral patterns that monitoring systems can detect to stop an operation from progressing.
Controls
The response cannot stop at the company's own perimeter. The trust relationships linking vendors, applications, identities, data and infrastructure need to be mapped:
- Inventory external services and the permissions granted to each, and revoke access for integrations no longer in use.
- Limit and monitor application credentials to the minimum necessary, rotating them immediately when exposure is suspected. Google and Mandiant highlight short-lived credentials and isolating the control plane.
- Check third-party code running in your environments, such as advertising, analytics, personalization, testing, customer service and payment scripts, and detect unexpected changes to those files with the same rigor applied to your own code.
- Segment networks, isolate administrative systems and maintain continuous behavioral monitoring, with the ability to revoke compromised credentials quickly.
- Assess vendors. In July 2026, NIST published a due diligence guide for supply-chain risk management, which recommends examining provenance, resilience, foundational security practices, ownership and control structure, and the different tiers of suppliers.
AI will also have to be built into defensive structures. Agents can correlate alerts, analyze logs, prioritize vulnerabilities and investigate suspicious activity, and automated systems can carry out pre-authorized responses, shortening the interval between identifying a threat and containing it. This creates a governance challenge: agents with access to credentials, administrative tools and critical infrastructure must operate under strict controls, such as least privilege, identity management, monitoring and human oversight.
The core guidance is to drop the idea that an organization's security can be assessed in isolation. A company can protect its own systems well and still depend on a vendor with privileged access, an application that runs external code or an integration that stays active for no reason. The question is no longer only whether the company is protected. It is whether it knows, monitors and can control the trust relationships it depends on.
The geopolitical backdrop
Once the facts are laid out, the question remains of who exploits these dependencies, to what ends, and how the interests of states, espionage groups and criminal organizations intersect in this environment. The documented cases themselves rule out a simple answer.
The same framework serves different goals. Check Point documented financial crime and digital certificate fraud operations. Microsoft links the ConnectWise campaign to Chinese state interests. Google saw signs of in-house frameworks belonging to espionage-linked groups, with a possible Chinese connection, and the investigation itself urges caution against turning suspicion of state ties into a definitive conclusion. State attributions, in the cases followed, come from private companies and carry caveats. Nothing in the material examined links all the groups to a single state.
The targets matter as much as the perpetrators. The FBI lists defense, semiconductors, energy and telecommunications, as well as financial services, government and health care. Google saw CAIRN in intrusions against U.S. aerospace organizations and an actor using PentAGI to scan organizations in Asia. The attacks on South Korean financial institutions exposed data on about 67,000 people. These are the sectors where countries' economic, military and technological advantage is concentrated.
The predominance of Chinese models such as DeepSeek and Kimi in the instances Check Point examined points to an economic and technical fact before a political one: these are cheap and accessible models for this use, and that does not prove state coordination. But it shows that dependence on models, like dependence on vendors, crosses borders and jurisdictions, and that control over tools, permissions and safeguards is spread among companies and countries.
The supply chain takes the question into the realm of sovereignty. Miljödata held data on about 2.2 million people, from Swedish municipalities, regions and government agencies. The ConnectWise campaign tried to reach more than 2,250 customer environments. When governments and essential services depend on a handful of vendors, a contest between states and criminal groups can be fought inside the infrastructure of people who do not even know they are in play.
The machine is not the model, it is the architecture
The episodes of October and November 2026 point to two mutually reinforcing shifts. The first is offensive: models, tools and agents organized into architectures that execute, evaluate and keep working, with human involvement reduced to oversight. The second is structural: an organization's security boundary no longer coincides with the boundary of its infrastructure.
The attack machine is not a single AI model. It is an architecture that turns models, tools and agents into operational capability, and that can be reproduced, scaled and distributed. But there is no proof that every intrusion is AI-driven. ARTEX is the documented case; in the public information consulted, Qubole and Optimizely have no equivalent evidence. AI did not create the fragility of the supply chain, but it can speed up its exploitation. And automation makes mistakes, leaves traces and depends on infrastructure, credentials and vulnerabilities that still have to exist.
The same logic serves defense. Research such as OpenAI's shows that finding and validating vulnerabilities can also be done at scale, and defense needs structures that monitor and respond continuously. When vendors share credentials, permissions, code and data access, a localized failure has consequences across a much larger network, and the time to identify and stop an exploit may be an organization's scarcest resource.
The political dimension remains. When the same tool serves state espionage and financial crime, and attribution comes from private companies with varying degrees of confidence, the response is no longer purely technical: it involves who can be held accountable, who sets the rules for the vendors that sustain governments and essential services, and how much each country depends on technology it does not control.
The question is no longer just which model can find a vulnerability. It is which architecture can turn that discovery into an operation, and which defensive architecture will be able to stop it before it has consequences.