FBI warns hackers weaponizing ARTEX AI framework to automate attacks on energy, manufacturing, telecom

Industrial Cyber

Source published: Site published:

Original article · فارسی

FBI warns hackers weaponizing ARTEX AI framework to automate attacks on energy, manufacturing, telecom

On November 17, 2026, the Federal Bureau of Investigation (FBI) warned in a FLASH advisory that hackers are using ARTEX, an open-source framework powered by large language models (LLMs), to automate many stages of cyberattacks. These stages include gathering information about targets, gaining initial access, increasing their system privileges, and stealing data. The FBI said malicious hackers have been using the framework since at least August to attack organizations in financial services, government, technology, defense, semiconductors, energy, healthcare, manufacturing, telecommunications, and legal services. Some attacks moved from the first stage of targeting to privileged access in less than three hours.

The FBI highlighted attacks against South Korean financial institutions that led to the theft of personal information belonging to about 67,000 people. It also described a supply chain attack that affected ConnectWise Cloud Backup customers and gave attackers access to more than 2,250 other customer environments. ARTEX uses specialized AI agents to turn broad goals into computer commands. This means attackers need very little human involvement after giving the system its first instructions.

ARTEX is an automated penetration testing platform built in the Go programming language. It uses several AI agents working together to turn high-level goals written in everyday language into computer commands. These commands are carried out by specialized superintelligence (SI) agents, so people need to do little beyond setting the original goal. The original developer removed the GitHub repository in October after reports of malicious use. However, ARTEX is still available through community-created versions, archived copies, local versions, and a commercial product.

Researchers found more than 1,600 ARTEX console instances that were publicly accessible in early November. The actual number of installations is likely higher because some systems are private or protected by firewalls.

The FBI warned that ARTEX makes complex cyberattacks easier, helps attackers target many organizations more quickly, and finds systems that security teams may overlook. These include supplier websites, older services, connections to third-party systems, and cloud infrastructure. The agency urged organizations to reduce the number of systems exposed to the internet, improve identity and access security, watch application activity, and keep evidence of suspected attacks.

ARTEX does not include its own SI model. Instead, it connects to outside LLM providers. These include advanced models such as Anthropic's Claude and OpenAI's GPT, as well as open-source or Chinese-developed models such as DeepSeek, GLM, Grok, and Kimi. An analysis of ARTEX systems controlled by attackers found that Chinese models were used most often. The DeepSeek Flash family was especially popular because it offers a balance between performance and low cost. The analysis also found no visible safety restrictions when requests were presented as penetration testing tasks.

The developer, who uses the name 'Autumn-27,' first released ARTEX on GitHub in July 2026. It quickly became popular, reaching an estimated 1,500 to 2,000 stars in about ten weeks. The developer said the tool was designed for authorized security testing within clearly defined limits. In early October 2026, the developer took the repository offline after reports that hackers were misusing it. However, this has done little to stop access because ARTEX remains widely available through independent copies and mirrors on code-sharing platforms.

Based on what researchers have observed and how attackers are adopting the tool, ARTEX creates several different risks. The first is that attackers can target many organizations at once and work faster. Its superintelligence agents automate the slow work of finding systems and exploiting weaknesses. This allows even small groups of attackers to run several campaigns at the same time and reach a scale that was previously beyond their abilities. The second risk is that attackers can find and abuse systems that organizations do not watch closely. ARTEX is particularly effective at finding less-protected systems outside an organization's main applications. These include partner and supplier websites, mobile application backends, employee applications, sales-support systems, older services, third-party connections, and applications built by different business units.

A third risk is supply chain attacks. The agents have already shown that they can find and exploit weaknesses in supply chains. One example involved changing the systems used to distribute iOS applications by using exposed cloud credentials. Because automated agents can map the trust relationships between organizations, attackers can use ARTEX to break into a less-protected third-party supplier, a CI/CD pipeline, or a software provider connected to another company. They can then move from that supplier into larger business networks that contain more valuable systems and data.

The tool also makes it easier for people with limited technical skills to carry out complex attacks. Finally, it shortens the time defenders have to detect and stop an attack. ARTEX can create large amounts of unusual network activity that may trigger security alerts. However, its speed can leave defenders with very little time to organize a response before attackers achieve their goals.

The FBI recommends that organizations check and investigate indicators of compromise (IOCs) before taking action, such as blocking an IP address. The agency points out that attackers may use cloud or virtual servers where IP addresses can change or be assigned temporarily. An IP address linked to malicious activity at one time may later be assigned to a harmless service. For this reason, these indicators should be treated as evidence of infrastructure that was used during a particular period. Organizations should confirm them using current network records or other intelligence sources before using them for security or investigations.

The FBI said that anyone reporting an incident involving ARTEX or a similar agent-based SI framework should provide a wide range of evidence. This should begin with information about the infrastructure involved. It includes source code or screenshots of suspected ARTEX console screens, along with any ARTEX configuration files.

Authentication and access records are also important. These should include service principal sign-in logs showing unusual source networks, autonomous system numbers (ASNs), hosting providers, or locations. Reports should also include user-agent strings from suspicious login attempts, IP addresses and login times, and evidence of logins from infrastructure that has not been seen before. Organizations should provide application and API activity records too. Examples include Microsoft Graph activity, Microsoft Entra audit records, and Microsoft 365 audit records that show signs of information gathering. Useful evidence includes a series of lookups performed one after another, unusually large numbers of records accessed during a single session on partner or employee websites, rate-limit violations, and unusually high volumes of API requests.

Information about the LLM and SI infrastructure used in an attack is also valuable. This includes outgoing connections to LLM API services, evidence of privately hosted LLM gateways or reseller tools such as CLIProxyAPI and New API, unfamiliar LLM proxy services or resellers in outgoing network logs, and unusual patterns in token use or API bills.

For cloud systems, reports should include AWS IAM activity logs, Azure ARM deployment validation dry-runs involving several types of resources, and test operations that write and delete cloud storage blobs within less than a second. Reports should also include evidence of stolen Kudu or SCM credentials, unusual Kudu API activity, container management logs, and queries to cloud metadata services. Unexpected uses of IAM roles, PassRole operations, and changes to instance profiles should also be reported.

Evidence of attacks and attempts to maintain access is important too. This includes web shells, BLOODALCHEMY components, signs of open-source remote access tools (RATs) such as WISP, vshell, and AdaptixC2, newly created accounts that follow unusual naming patterns, and changes to credentials used by OAuth applications. Reports should also include new SSH keys, Elastic IP address allocations, and unexpected launches of EC2 or Azure virtual machines. Organizations should describe how the attack happened, including short bursts of activity, large numbers of failed attempts that change gradually, multiple sub-agents overwriting one another or creating duplicate requests, and large numbers of notifications or emails sent from compromised applications.

Finally, reports should explain the damage and impact. This includes the total number of compromised accounts or affected users, the types and amounts of data accessed or stolen, the time between the first suspicious activity and its detection, the time between detection and containment, and the total amount of attempted financial losses.

The FBI's main immediate recommendation is for organizations to reduce the number of systems they expose to the internet to protect against ARTEX and similar agent-based SI frameworks. Organizations should start by listing every application that can be reached from the internet. They should pay special attention to less-visible systems such as partner and supplier websites, mobile application backends, employee applications, sales-support systems, older services, and third-party connections. Sensitive systems, including cloud administration tools, network device management systems, Kudu and SCM publishing interfaces, hypervisors, and backup platforms, should only be accessible through trusted networks or dedicated administrative connections.

Basic publishing authentication should be turned off when it is not needed. Organizations should consider using run-from-package deployment methods. They should also require an architecture review before giving any workload or management interface a public IP address. Where possible, enterprise applications should be placed behind Zero Trust access systems so that they are no longer directly exposed to the internet.

For identity and application security, the guidance recommends giving multi-tenant application permissions and consent grants only the access they actually need. Organizations should closely watch for new credentials, changes to privileges, and unusual activity by applications acting without a signed-in user. Organizations that have ever installed or approved the affected ConnectWise applications should review every Microsoft 365 tenant involved. They should search using application IDs rather than application display names. They should also check successful and failed service principal sign-ins separately instead of relying only on the IP addresses published by the vendor.

Organizations should review Microsoft Entra audit records for newly added credentials, service principals, role assignments, and consent grants. If suspicious activity is confirmed, they should remove the affected permissions and change the related credentials. Business APIs should have protections such as rate limits and transaction limits. Websites should alert security teams when users perform a series of lookups or access unusually large numbers of records. Applications should also avoid showing system error details or internal information to people who are not signed in. Security teams should check configurations, development pipelines, code repositories, and container images for passwords or other secrets stored inside them. Any exposed secrets should be changed.

If an attack is suspected, organizations must preserve evidence while still taking urgent steps to contain the threat. This includes Microsoft Entra sign-in and audit records, Purview Audit records, security information and event management (SIEM) data, and any Microsoft Graph activity records that are already being collected. Reviews of activity connected to ConnectWise should go back at least to October 1, 2026, and further if there are signs of earlier suspicious activity. Only properly authorized incident responders should remove confirmed unauthorized credentials, permission grants, owners, and identities. They should do this only after collecting the evidence. Affected ConnectWise application connections should remain disabled until the vendor provides official instructions for restoring them. Organizations should not delete the enterprise applications as their first response, because doing so could destroy important settings and evidence.