The Federal Bureau of Investigation (FBI) warned in a FLASH advisory that hackers are using ARTEX, an open-source, large language model (LLM)-driven autonomous penetration testing framework, to automate multiple stages of cyber intrusions, from reconnaissance and initial access to privilege escalation and data exfiltration. The agency said malicious cyber actors have used the framework since at least August to target organizations across financial services, government, technology, defense, semiconductors, energy, healthcare, manufacturing, telecommunications and legal sectors. Some intrusions progressed from initial targeting to privileged access in less than three hours.
The FBI highlighted campaigns targeting South Korean financial institutions that resulted in the theft of personal information belonging to approximately 67,000 individuals, as well as a supply chain compromise affecting ConnectWise Cloud Backup customers and enabling access to more than 2,250 downstream tenants. The framework uses specialized AI agents to translate high-level objectives into machine-executable commands, requiring minimal human intervention beyond initial instructions.
ARTEX is an autonomous, multi-agent penetration testing platform written in Go. It translates high-level tactical goals, described in plain natural language, into machine commands carried out by specialized super intelligence (SI) agents, so little human involvement is needed beyond setting the initial objective. Although its original developer removed the GitHub repository in October following reports of malicious use, ARTEX remains accessible through community forks, archived copies, localized versions and a commercial offering.
Researchers identified more than 1,600 publicly exposed ARTEX console instances in early November, with deployments likely higher when private and firewall-protected systems are included.
The FBI warned that the framework lowers technical barrier to complex attacks, accelerates multi-target operations and exploits overlooked systems, including supplier portals, legacy services, third-party integrations and cloud infrastructure. It urged organizations to reduce internet-facing attack surfaces, strengthen identity and access controls, monitor application activity and preserve evidence of suspected intrusions.
The ARTEX framework does not ship with its own SI model. Instead, it connects to external LLM providers, including frontier models such as Anthropic’s Claude and OpenAI’s GPT, as well as open-source or Chinese-origin models such as DeepSeek, GLM, Grok and Kimi. Analysis of attacker-controlled ARTEX instances shows that Chinese models are the most widely used, with the DeepSeek Flash family dominating because it balances capability with low cost and shows no observable guardrails when a request is framed as penetration testing.
The developer, who goes by ‘Autumn-27,’ first released ARTEX on GitHub in July 2026, and it gained rapid adoption, reaching an estimated 1,500 to 2,000 stars within about ten weeks. The developer said the tool was meant for authorized security testing within defined scopes, and took the repository offline in early October 2026 after reports that hackers were abusing it. That move has done little to limit access, however, since ARTEX remains widely available through independent routes and mirrors on code-sharing platforms.
Based on observed capabilities and ongoing threat actor adoption, ARTEX presents several distinct risks. The first is accelerated multi-target operations. Its super intelligence agents automate the time-consuming phases of enumeration and exploitation, which allows even small threat groups to run concurrent intrusion campaigns at a scale that was previously out of their reach. The second is the exploitation of overlooked attack surfaces. ARTEX is particularly good at finding and abusing peripheral systems that receive less security attention than flagship applications, such as partner and supplier portals, mobile backends, employee applications, sales-support systems, legacy services, third-party integrations, and applications created across different business units.
A third risk is supply chain compromise. The agents have already shown that they can identify and exploit supply chain attack vectors, including modifying iOS application distribution infrastructure through exposed cloud credentials. Because autonomous agents can systematically map trust relationships between organizations, attackers can use ARTEX to compromise a weaker third-party vendor, CI/CD pipeline, or integrated software provider and then pivot from that peripheral supplier straight into high-value upstream enterprise networks.
The tool also lowers the barrier to entry, since threat actors with minimal expertise can now carry out complex and sophisticated campaigns. Finally, it compresses the detection window. ARTEX generates high volumes of forensically noisy traffic that may trigger alerts, but attack speed leaves defenders little time to mount a coordinated response before the objective is completed.
The FBI recommends organizations investigate or vet IOCs prior to taking action, such as blocking. It recognizes that infrastructure used for malicious activity may be hosted on cloud or virtualized platforms where IP addresses can be dynamically or temporarily assigned. As a result, an address associated with malicious activity at one point in time may later be reassigned to benign services. Accordingly, these indicators should be interpreted as historically observed infrastructure within the activity window and should be corroborated with current network telemetry or additional intelligence sources when used for defensive or investigative purposes.
The agency said that anyone reporting an incident involving ARTEX or a similar agentic SI framework is asked to provide a broad set of evidence, beginning with the infrastructure itself. This includes source code or screenshots of suspected ARTEX console interfaces, along with any ARTEX configuration files.
Authentication and access logs are also needed. These should include service principal sign-in logs that show unusual source networks, ASNs, hosting providers, or geographies, together with the user agent strings from suspicious authentication attempts, the IP login details and their timestamps, and any sign of authentication from previously unseen infrastructure. Reporters should add records of application and API activity as well, such as Microsoft Graph activity, Microsoft Entra audit records, and Microsoft 365 audit records that show reconnaissance. Sequential lookups, unusually high numbers of records accessed per session on partner or employee portals, rate limit violations, and volumetric anomalies in API calls are all worth including.
Indicators of the LLM and SI infrastructure behind an attack are equally valuable. These include outbound connections to LLM API endpoints, evidence of self-hosted LLM gateways or reseller software such as CLIProxyAPI and New API, unfamiliar LLM proxies or resellers in egress logs, and unusual token consumption or API billing patterns.
On the cloud side, reporters should share AWS IAM activity logs, Azure ARM deployment validation dry-runs across multiple resource types, and test blob write and delete operations in cloud storage with sub-second timing. Evidence of compromised Kudu or SCM credentials, unusual Kudu API activity, container orchestration logs, and cloud metadata service queries should be included, as should unexpected IAM role assumptions, PassRole operations, and instance profile modifications.
Artifacts of exploitation and persistence matter too. These cover web shells, BLOODALCHEMY components, open-source RAT indicators such as WISP, vshell, and AdaptixC2, newly created accounts that follow naming patterns, and credential modifications to OAuth applications. SSH key additions, Elastic IP allocations, and unexpected EC2 or Azure VM launches belong here as well. Reporters should also describe the characteristics of the attack itself, such as compressed bursts of activity, high volumes of failed attempts that change iteratively and incrementally, multiple sub-agents overwriting each other or generating duplicate requests, and mass notification or email broadcasts from compromised applications.
Finally, the report should convey the loss and impact, including the total number of compromised accounts or affected users, the types and volumes of data accessed or exfiltrated, the time from initial suspicious activity to detection, the time from detection to containment, and the total attempted loss amounts.
The FBI’s immediate priority for organizations is to shrink their external attack surface in response to ARTEX and similar agentic SI frameworks. That begins with inventorying every internet-facing application, especially peripheral ones such as partner and supplier portals, mobile backends, employee applications, sales-support systems, legacy services, and third-party integrations. Sensitive services, including cloud administration, network device management, Kudu and SCM publishing endpoints, hypervisors, and backup platforms, should be confined to trusted networks or dedicated administrative paths.
Basic publishing authentication should be disabled where it is not needed, run-from-package deployment models are worth considering, and architecture review should be required before any workload or management interface is given a public IP. Where feasible, enterprise applications should sit behind Zero Trust access solutions so they are no longer exposed externally at all.
On identity and application protection, the guidance calls for least privilege on multi-tenant application permissions and consent grants, along with close monitoring of credential additions, privilege assignments, and unusual app-only activity. Organizations that have ever installed or authorized the affected ConnectWise applications should review every Microsoft 365 tenant involved, searching by application ID rather than display name, and should examine successful and failed service-principal sign-ins separately instead of trusting published vendor IPs alone.
Microsoft Entra audit records deserve a look for new credentials, service principals, role assignments, and consent grants, and confirmed suspicious activity should lead to revoked permissions and rotated credentials. Business APIs need guardrails such as rate limits and transaction thresholds, portals should alert on sequential lookups and unusually large record access, and applications should avoid exposing stack traces or internal details to unauthenticated users. Teams should also scan configurations, pipelines, repositories, and container images for embedded secrets and rotate any that are exposed.
When an incident is suspected, evidence must be preserved without delaying urgent containment, including Microsoft Entra sign-in and audit records, Purview Audit records, SIEM data, and any Microsoft Graph telemetry already configured. Reviews of ConnectWise-related activity should reach back to at least 1 October 2026, and further if earlier anomalies appear. Only appropriately authorized responders should remove confirmed unauthorized credentials, grants, owners, and identities, and only after evidence has been captured. Affected ConnectWise application connections should stay disabled until the vendor issues formal restoration instructions, and the enterprise applications themselves should not be deleted as a first response, so that configuration and evidence remain intact.


