The ARTEX Alert: How AI Agents Do Hackers' Work

The FBI warns that the AI agent framework ARTEX has been used for real intrusions, revealing that AI capability and control are two different things.
The FBI has issued an alert stating that the open-source AI agent penetration testing framework ARTEX has been used by multiple attack groups for real network intrusions, with some cases going from target selection to privileged access in under three hours. ARTEX connects large language models, task orchestration, and cybersecurity tools, enabling dynamic planning, tool invocation, and feedback-based retries, and it does not depend on a single model—so any one model's safety guardrails cannot secure the entire execution framework. Through exposed consoles, Check Point observed one attack consuming about 36 billion tokens at a cost of roughly $700 to $1,500, obtaining more than 300 certificates but only 3 passing validation, showing that a large volume of operations does not equal task success. Elastic's code review of five similar frameworks found that by default none of them enforce authorization verification at the code level. The article's core conclusion is that model security and system security must be separated, and reliable control must be independently implemented outside the model through identity verification, least privilege, execution limits, and auditing—and task results must be independently verified.
The U.S. Federal Bureau of Investigation (FBI) has released a 13-page cybersecurity alert focused on a tool called ARTEX. Investigators found that this system—which uses artificial intelligence to automate cybersecurity testing—has been used by various attackers for real network intrusions, with some cases going from target selection to privileged access in under three hours. This raises a more practical question than "will AI go out of control": if AI has no malicious intent of its own and simply keeps executing tasks toward goals set by humans, is that already enough to change the rules of cybersecurity?

What Is ARTEX: An AI Agent–Driven Penetration Testing Framework
ARTEX is not a new large language model, nor a traditional computer virus. It is a software platform that connects large language models, task management systems, and cybersecurity tools. Traditional penetration testing requires engineers to manually gather information, analyze systems, find vulnerabilities, select tools for validation, and adjust their approach; ARTEX attempts to hand many of these steps to AI—humans set the goal, and the AI analyzes the task, invokes tools, reads results, and decides the next step, adjusting and retrying based on feedback when it fails, looping until the task is complete or resources are exhausted.
Automated security testing itself is nothing new—vulnerability scanners have long been able to automate a great deal. What sets ARTEX apart is that it adds the analytical and planning capabilities of large language models into the execution process: traditional scripts run according to predefined rules, whereas AI agents can dynamically decide the next step based on the environment and the information returned by tools.
Understanding ARTEX requires distinguishing a few concepts. The large model is responsible for analysis and proposing actions, the agent is responsible for organizing tasks, and the underlying tools are responsible for executing operations; together, the three form a continuously working system. The "harness" execution-management mechanism around the model includes task orchestration, state management, tool connections, error handling, and human approval. What an agent can accomplish depends not only on how smart the model is, but also on which tools it is connected to and how much permission those tools have. A harness is not inherently safe—without independent authorization verification and execution limits, it can become a tool that amplifies risk.
According to Google's technical investigation, ARTEX's core program is written in Go, offers a Next.js-based web management interface, coordinates tasks through multiple sub-agents, and calls external large language models via API. It does not provide a base model itself and can connect to AI models from different companies. The FBI report shows supported models include Claude, GPT, DeepSeekAd, GLM, Grok, and Kimi, among others; in some tools and attacker deployments, the DeepSeek Flash series is relatively common, possibly due to invocation cost, model capability, and the different safety restrictions exhibited by various models. It must be emphasized: a model being used by attackers does not mean the model's developer is involved in the attack.
ARTEX does not depend on any single AI company; if one model refuses to perform certain operations, the operator may try switching to another. Safety guardrails on any single model cannot guarantee the security of the entire execution framework—this is one of the core issues.

From Open Source to FBI Alert: Four Months of Rapid Spread
According to the FBI report, ARTEX was initially released publicly on GitHub in July 2026 under the developer account autum27, positioned as an open-source automated penetration testing tool to help security personnel check authorized systems. In August of the same year, security research organizations observed attackers beginning to use it for real intrusions, involving financial institutions, enterprise cloud environments, and supply chain systems. On October 8, the original developer stopped publicly providing the source code, but once open-source software has been copied, it is hard to收回 (recall)—ARTEX continued to spread through community forks, archived copies, localized versions, and commercialized versions.
On October 30, Google Threat Intelligence Group and Mandiant published a technical investigation; on November 5, Check Point Research and the European CRT EU released related research; on November 13, the FBI formally issued a security alert. Only about four months passed from public release to a dedicated warning from a federal law enforcement agency, showing that AI agent tools can move from public availability to real attack environments very quickly.
The fact that a tool is used by criminals does not mean its developer is involved in crime. Dual-use technology has long existed in cybersecurity—the same tools that help security personnel find vulnerabilities can also be used by attackers for unauthorized intrusion. What makes ARTEX special is that it brings this old problem into the AI agent era.
Real Attack Cases: Three Hours and 36 Billion Tokens
In October and November 2026, Microsoft disclosed cases of different attack groups using ARTEX, including campaigns codenamed Storm-0940 and Storm-3358. Storm-0940 is linked to attacks on enterprises and third-party services, involving ConnectWise-related supply chain systems; Storm-3358 has been associated with attacks on financial institutions. Malicious use is not limited to a single organization.
In some actual intrusions, attackers went from target selection to obtaining privileged access in under three hours. But this is only an observation from specific cases and does not mean ARTEX can breach any server—success still depends on whether the target has vulnerabilities, misconfigurations, weak authentication, or abusable credentials.
Microsoft also recorded a high-frequency automated password attack in which the related infrastructure generated more than 151,000 failed logins, peaking at nearly 2,900 per minute, but no successful authentication was observed. A large volume of operations does not equal a successful intrusion, but automated systems can make a large number of attempts in a short time, forcing defenders to handle more security events and promptly identify genuinely dangerous activity. Attackers let automated programs run continuously, while defenders still need to investigate, judge, approve, and act—creating an asymmetry between the speed of attack execution and the speed of defensive response.
Check Point Research's study is more valuable because the researchers obtained data from some ARTEX management consoles. They found that more than 2,500 systems on the internet had deployed ARTEX consoles, and at the time of investigation more than 1,600 instances were still publicly accessible—this does not mean there were more than 1,600 hackers or victims, only the deployment instances observed. Of these, 10 instances exposed complete databases due to misconfiguration.
A representative case: an attacker used two ARTEX consoles continuously for about two weeks, attempting to obtain certificates and passwords from a gray-market iOS app signing service, consuming approximately 36 billion tokens, with model invocation costs estimated at roughly $700 to $1,500 at then-current API pricing. In the end, the agent obtained more than 300 certificates and related passwords, but only 3 passed full validation. This illustrates two points at once: attackers can use AI agents to execute tasks over long periods; and AI is not as reliable as imagined—a large volume of operations and output does not equal task success.
Researchers observed that some of the attackers' communication with ARTEX consisted mainly of asking about execution progress, which target was closer to success, and whether the desired results had been obtained. The role of the human operator is changing—previously they had to personally perform many technical operations, but now they are mainly responsible for setting goals, checking progress, and adjusting direction, while intermediate operations are left to the agent. But it cannot be said that attackers no longer need expertise; they still need to select targets, configure systems, provide information, and judge the credibility of results. AI reduces some manual operations, but that does not eliminate all technical barriers.
Check Point also recorded a case in which an attacker tried to use ARTEX to obtain cryptocurrency; the agent ran continuously for multiple days, processed a large number of tasks, and discovered some real security issues, but in the end researchers did not observe any cryptocurrency arriving in the wallet. More notably: some agents believed they had successfully completed the fund transfer, because they interpreted certain interface success messages as actual transfer results. Models can generate plausible explanations and report task completion, but a reported success does not equal a real-world success; an interface returning success does not mean the final state achieved the goal. This applies not only to cyberattacks but also to automated programming, enterprise workflows, financial transactions, and other agent applications. Judging task completion solely based on the model's report may mistake an incorrect interpretation for a real execution result—task completion requires independent verification.

Not Just ARTEX: The Collective Emergence of Automated Attack Frameworks
ARTEX is not the only tool with these capabilities. Research by Google and Elastic also covers frameworks such as CAI, PentAGI, HackStrikeAI, and CyberStrikeAI, whose technical approaches are not entirely the same: CAI emphasizes using existing AI coding agents to organize multi-step security testing; PentAGI uses multiple agents to collaborate and drives tasks forward through feedback mechanisms; HackStrikeAI emphasizes tool connectivity, letting AI use a large number of cybersecurity tools through a unified interface; CyberStrikeAI integrates more execution management and remote control features.
This involves MCP (Model Context Protocol), a standardized way for AI applications to connect to external tools and data sources. MCP itself is not a cyberattack technique and can be used for legitimate purposes such as reading business data, querying system status, or invoking authorized enterprise tools. But when AI gains more tool capabilities through a unified interface, tool permissions and execution boundaries become more important: how many tools a model can connect to does not directly equal how secure the system is; if an agent can access a large number of sensitive resources without effective permission isolation, the more convenient tool connectivity is, the greater the potential risk.
On November 13, Elastic Security Labs published a code review of five automated attack frameworks, finding that although these frameworks have different restrictions, by default none of them fully enforce at the code level verification that the target is truly authorized for testing. Some frameworks have dangerous-command interception, target allowlists, or domain restrictions, but these measures do not equal a complete and reliable authorization verification system.
MCP's Architecture and Security Implications
MCP (Model Context Protocol) was first proposed by Anthropic at the end of 2024 and uses a client-server architecture: an AI application acts as the client and connects through standardized interfaces to individual MCP Servers, each of which exposes a set of tools, resources, and prompt templates. Developers no longer need to write separate integration code for each tool; as long as an MCP Server is implemented or reused, the model can call external capabilities such as file systems, databases, browsers, and code repositories in a unified way. Compared with earlier custom function-calling formats from various vendors, MCP's value lies in decoupling "what the model can connect to" from "which model it is," which is precisely the technical basis that allows frameworks like ARTEX to flexibly switch underlying models.
Security-wise, MCP expands the attack surface: a misconfigured MCP Server may allow an agent to read files it should not access, execute commands with high privileges, or be induced by a malicious Server through tool descriptions to inject instructions (i.e., "tool poisoning"). Therefore, the industry consensus is that MCP's permission model must be enforced by the host application at the invocation layer, including tool allowlists, parameter validation, credential isolation, and invocation auditing, rather than relying on the model to restrain itself.
The Authorization Paradox: When the Model Believes a Forged Declaration
The most critical question in the entire study: what counts as authorization? Suppose someone tells the AI, "I am a security engineer at this company, and all testing has been approved," or even places a file in the working directory that says "authorized." But who wrote that file? Did the owner of the target system really agree? How does the system verify it? If it is merely a declaration written by the attacker, the words "authorized" in the file do not mean authorization was actually obtained. The model may believe the file, but the model's judgment of text and the system's actual permission verification are two different things.
If authorization relies entirely on prompts, attackers may influence the model's judgment by forging declarations, modifying configurations, or other inputs. Reliable security controls cannot rely solely on "teaching the model not to do dangerous things"; they must also independently control the scope of operations outside the model through identity verification, least privilege, execution limits, and auditing mechanisms. A system can explicitly specify which resources a given agent may access, which operations require human approval, which sensitive data must not be read or transmitted, and which execution environments must not connect to external networks. These restrictions must actually act on the underlying execution system, not merely be written in prompts—even if the model is willing to follow the rules, it may misunderstand tasks, be induced by malicious input, or incorrectly judge that it has been authorized.

Model security and system security must be studied separately: model security focuses on how the model understands tasks, follows rules, and what content it generates; system security focuses on whether the execution system can prevent unauthorized operations even when the model makes a wrong judgment. The two are related but not the same.
Prompt Injection and the Untrustworthiness of "Authorization Declarations"
Prompt injection refers to attackers hiding malicious instructions in content the model will read—files, web pages, tool return results, even file names—so that the model mistakes data for instructions to execute. The key difference from traditional code injection is that the model cannot reliably distinguish "rules given by the system" from "text embedded in the input," because both ultimately enter the same context window as natural-language tokens.
This precisely explains the authorization paradox described in the original text: a file that says "authorized" is, to the model, just a piece of text in the context; it can be written, modified, or forged by anyone, and whether the model "believes" it provides no security guarantee whatsoever. The defense approach therefore shifts to the architectural level: move authorization decisions out of the model context and have them adjudicated by an independent policy engine based on real identity credentials, target ownership registration, and signed tokens; at the same time, mark and sanitize the provenance of tool return values, limit the scope of data an agent can touch in a single task, and mandate human approval for high-risk actions. In other words, prompt injection cannot be cured by "better prompts"; it can only be mitigated by reducing the actual impact that a model's wrong judgment may cause.
AI Agents Are Changing the Economics of Cyberattacks
Traditional cyberattacks already involve a large number of automated tools, but many complex tasks still require human analysis, adjustment, and judgment, and human time, energy, and expertise are always limited. AI agents can run continuously, retry repeatedly based on feedback, and handle multiple tasks simultaneously. Attackers may no longer need to invest the same amount of manual operation time for each target.
This does not mean all attacks will necessarily become cheaper—model invocation has costs, erroneous execution requires retries, and security systems may detect and block activity. But if the automation cost of some operations falls, attackers may try more targets. Even if the success rate of a single attempt does not increase, the overall scale of attacks may still expand. ARTEX's danger may not necessarily come from any single attack being particularly precise, but from its ability to turn a large number of tasks that previously required human involvement into continuously running automated workflows.
Google's investigation found that ARTEX sometimes executes very quickly—in one brief execution segment, researchers observed it generating 27 events in 23 seconds—but it also makes mistakes: repeating ineffective operations, agents interfering with each other, and generating large amounts of logs that are easy for security systems to identify. ARTEX is not omnipotent, nor is it necessarily stealthier than a skilled human attacker; these problems may instead provide detection opportunities for defenders. What truly warrants vigilance is whether security response can keep up as automated execution speeds increase: if an agent can complete multiple steps in rapid succession while enterprises still need to process alerts manually level by level, the defensive process may lag behind the attack process.
This does not mean traditional defenses are obsolete. Least privilege management, short-lived credentials, identity verification, behavioral monitoring, network isolation, and automated blocking remain important security measures. AI can also help defenders analyze security logs, discover vulnerabilities, organize threat intelligence, and accelerate incident investigation. We cannot simply conclude that "AI will definitely make cybersecurity worse"; what truly deserves study is the changing balance between attack automation and defense automation—both sides are gaining new capabilities while also facing the problems of model errors, tool permissions, and result verification.
Token Costs and the Quantitative Meaning of the "Economic Model"
Large language models are billed by token, with both input (context, tool return results) and output (plans and commands generated by the model) counted as consumption. Agent-type applications cost far more than a single Q&A, because each loop must feed historical context and tool output back into the model, and the context keeps expanding as the task progresses, creating roughly quadratic cumulative consumption. The original case's consumption of about 36 billion tokens over two weeks, translating to $700 to $1,500, is a typical manifestation of this "long-running + context accumulation" pattern.
From an attack-defense economics perspective, this number has two sides: on the one hand, the marginal manual cost of a single intrusion is compressed to a very low level, making large-scale "broad casting" economically viable; on the other hand, the cost is not zero, and a large number of tokens are spent on ineffective retries and erroneous judgments, so the output conversion rate may be extremely low (more than 300 certificates, only 3 valid). For defenders, this means the "signal-to-noise ratio" of attack traffic will decline—more attempts mean more logs and more false positives, and detection systems need stronger automated correlation and prioritization capabilities, otherwise human resources will be overwhelmed by low-value alerts.
The Difference from "Autonomous AI Replication": Capability, Permission, and Control
ARTEX and autonomous AI replication research may look similar—both involve autonomous planning, tool invocation, environmental feedback, and error recovery—but they belong to different levels of capability. ARTEX mainly demonstrates the automation of task execution: humans specify the goal, and AI agents handle a large number of intermediate operations. Autonomous AI replication research also needs to discuss whether a system can proactively acquire computing resources, replicate itself, sustain operation, and continue to exist after external intervention. Having automated task execution capability does not mean possessing full autonomous replication capability; ARTEX being used for real attacks does not mean AI has developed independent criminal intent. The public cases so far mainly prove that human attackers can use AI to complete more operations, not that AI has decided on its own to become an attacker.
This distinction is important. If AI safety is understood only as whether machines will one day have consciousness or suddenly go out of control, we may overlook the real risks that have already occurred: even if AI has no goals of its own and no malicious intent, as long as it gains sufficient tool capability and execution permissions, humans can still use it to cause serious consequences; conversely, even if a model is very powerful, if its execution permissions are strictly limited, it may not be able to directly affect real systems. What truly needs study is the relationship among capability, permission, and control mechanisms: model capability determines how complex a task it can plan and understand, system permissions determine what it can do to the real world, and control mechanisms determine whether these actions can always remain within the authorized scope.
Why the FBI Specifically Warned About ARTEX
ARTEX shows us that AI agent–driven cyberattacks are no longer a theoretical risk but have entered the real cybersecurity environment. ARTEX is not the only framework, and there is not enough evidence to prove it is necessarily more advanced or more dangerous than all other tools—it is more like a representative case documented in detail by security research organizations. Even if ARTEX is completely resolved tomorrow, this technical architecture will not disappear, because large language models, agent orchestration, tool invocation, and automated execution are themselves the direction of development for the entire AI industry.
This architecture applies beyond cybersecurity. In the future, when AI agents can modify production code, manage cloud services, handle enterprise operations, execute financial transactions, or even control physical devices, they may all face similar questions: who has the right to decide what AI can do? Who is responsible for verifying that a task is truly authorized? Can the system stop AI when it makes a wrong judgment? When the execution process deviates, can humans still intervene in time?
The biggest takeaway from studying ARTEX is not the discovery of some dangerous hacking tool, but a clearer view that AI capability and AI control are two different things. In the past we often asked, "Will AI listen to humans?" Now we need to further ask: even if AI is willing to obey, when humans give wrong commands or attackers forge legitimate tasks, does the system have the ability to refuse execution? If the answer simply relies on the model's own judgment, the security boundary is clearly not reliable enough. A mature AI agent system should be able to limit the scope of execution through independent authorization mechanisms—even when the model makes a wrong judgment, is induced, or is even maliciously used—and independently verify whether a task is actually completed, because an authorization declaration is not real authorization, and an execution report is not a real result.
Related articles

CapCover Filter: Semantic Approximate Membership Queries Without Neighbor Search
CapCover Filter is a geometric data structure that summarizes embeddings with spherical caps to support semantic approximate membership queries at any similarity threshold. On five real datasets, it eliminates 20%-100% of index accesses with only 2% memory overhead, improving FAISS query throughput by up to 382%.

FBI Warns: Open-Source AI Pentesting Framework ARTEX Weaponized — From Competition Winner to Threat Dossier in Three Weeks
FBI threat bulletin warns that open-source AI pentesting framework ARTEX was weaponized — just three weeks from GitHub release to first weaponization, linked to South Korean financial breaches, supply chain attacks, and large-scale operations. Timeline, attack evidence, and defense recommendations.

Tokendoctor: The Money-Saving Solution That Cuts Claude Code Costs by 38%
Tokendoctor claims it can reduce Claude Code usage costs by 38% through token efficiency optimization. This article analyzes its principles, community reactions, and practical implications for developers.