Skip to contentFreeCyberSocialCon 2026Tue, Dec 8 – Thu, Dec 10 · Online · Live talks on DFIR, OSINT and digital forensicsSee the agenda →

ARTEX: What the FBI's AI Pentest-Framework Warning Means for Investigators

The FBI says actors are using ARTEX, an open-source AI penetration testing framework, to move from targeting to privileged access in under three hours. Here is what it can do, what evidence it leaves, how to triage it, and how to protect against it.

Nov 17, 2026 · 6 min read
Dim digital forensics lab desk at dusk with a write-blocker, cables, evidence bags, a notebook and dark monitors

The first question in many intrusion cases is what the attacker did. With ARTEX, an examiner may also have to ask how much of it a person did. On November 13, 2026, the FBI published FLASH-20261113-001, a warning about ARTEX, an open-source, LLM-driven penetration testing framework that criminal and state-aligned actors have used to chain reconnaissance, exploitation and data theft. The FBI says some intrusions went from initial targeting to privileged access in under three hours. For investigators, that speed squeezes the evidence window: logs roll over and the activity arrives as dense bursts, not a human's pace.

What ARTEX is, and why it is in today's headlines

Established (FBI, vendors): ARTEX was released on GitHub by a developer using the name "Autumn-27" in July 2026, and it does not include its own model. It connects to external LLM providers, including frontier and open-source models (FBI). The developer says it was meant for authorized testing, and the FBI says the repository was made unavailable on October 8, 2026 after abuse was reported. TechCrunch describes this as the developer making the software closed-source. Forks, mirrors and a commercial version called ARTEX PRO keep it in circulation (FBI; Google Threat Intelligence Group).

The story peaked this week because the FBI consolidated vendor research into one advisory. It names two campaigns: theft of personal information on about 67,000 people at South Korean financial institutions, and a supply chain compromise of ConnectWise Cloud Backup that led to attempted sign-ins to more than 2,250 downstream Microsoft 365 tenants (FBI). TechCrunch reports the South Korea figure as at least 68,000 and notes that ConnectWise told it that it cannot confirm the number of affected environments cited by the FBI.

What ARTEX can do

Research finding: An operator supplies a target and an objective. A planner agent breaks the objective into tasks, and worker agents run them with ordinary tools, recording results on a shared graph (Datadog; Talos). Microsoft reports that defaults include Nmap, curl, wget, Python HTTP clients and headless Chromium (Microsoft). Capabilities seen in incident response include automated Active Directory reconnaissance, privilege escalation, harvesting of Azure storage keys, and generation and deployment of custom web shells (Google Threat Intelligence Group).

Documented tradecraft

  • Initial access: unsigned SAML assertions, XXE, SSRF, anonymous file upload, command and template injection, SQL injection, IDOR, missing API authentication, hardcoded administrator passwords in JavaScript bundles, and stolen VPN configurations (FBI).
  • Persistence: JSP and PHP web shells, added credentials on existing OAuth applications, and open-source RATs (WISP, vshell, AdaptixC2) downloaded with curl and netcat (FBI; Microsoft).
  • Cloud and identity abuse: Microsoft says Storm-0940 used ARTEX for password spraying and OAuth application abuse against downstream tenants, followed by Graph reconnaissance, application credential changes and attempted privilege changes (Microsoft). Datadog saw anonymous Cognito sign-ups and long-term access keys used to search CloudWatch logs for financial data and secrets (Datadog).
  • Overlooked systems: the FBI says ARTEX is especially effective against partner and supplier portals, mobile backends, legacy services and third-party integrations (FBI).

Microsoft states that the ConnectWise activity did not result from a vulnerability in Microsoft products. It also says that a successful authentication does not by itself show that data was accessed or exfiltrated (Microsoft).

What evidence ARTEX leaves behind

Established: The FBI says ARTEX produces "forensically noisy" traffic. Most of what follows comes from vendor observations, so treat it as a source of hunting hypotheses rather than a signature set.

Network and identity logs

  • Tempo and shape: compressed bursts, high volumes of failed attempts with small iterative changes, sequential enumeration of login and search endpoints, and duplicate requests when sub-agents overlap (FBI).
  • Probe habits: Datadog reports that ARTEX requests a path that should not exist, such as /zzz-nonexistent-xyz123 or an S3 key like zzz_nonexistent_abc.txt, to learn what "not found" looks like before a burst of guesses (Datadog).
  • User agents: strings such as artex-pentest/1.0, ARTEX-worker/1.0 and Mozilla/5.0 ARTeX appear in Datadog's indicator list. Operators can change them, and Microsoft notes that blocking such traffic will not stop an AI agent, which can quickly change it (Datadog; Microsoft).
  • Egress to LLM providers: outbound calls to LLM API endpoints, self-hosted gateways or resellers, and unusual token-consumption patterns are on the FBI's list of items to collect (FBI).

Cloud and host artifacts

  • CloudTrail: unauthenticated S3 reads (userIdentity.type of AWSAccount, accountId of anonymous), repeated FilterLogEvents calls with a limit of 1000, Cognito SignUp and InitiateAuth sequences, and bursts of AccessDenied errors during discovery (Datadog).
  • Azure and Entra: sub-second permission checks, ARM deployment dry-runs, test blob writes and deletes, new application credentials, and role assignments (FBI).
  • Hosts: web shells, newly created accounts, SSH key additions, Elastic IP allocations, unexpected VM launches, and RAT indicators (FBI).

How to triage and investigate

Practitioner interpretation: The advisory's guidance maps to a short order of operations for examiners and responders.

  1. Preserve before you clean. Export Entra sign-in and audit records, cloud control-plane logs and SIEM data first. The FBI says not to delay urgent containment while waiting for exports, and not to delete enterprise applications as a first response, because that destroys configuration evidence (FBI).
  2. Establish the window. For the ConnectWise activity, the FBI advises reviewing from at least October 1, 2026 and searching by application ID rather than display name (FBI).
  3. Separate sign-ins from changes. Review successful and failed service-principal sign-ins separately, then look for credential additions, new owners, consent grants and privileged role assignments (FBI).
  4. Do not stop at the published IP list. The FBI says to investigate unfamiliar sources outside expected vendor infrastructure, and that infrastructure indicators may be reassigned to benign services (FBI).
  5. Assume logging gaps matter. Barracuda said that its logs were not detailed enough to rule out access, so it chose to "assume the worst" and rotated customer credentials (TechCrunch).
  6. Document attribution carefully. Record timing and request patterns separately from conclusions about who operated the tool.

How to protect

  • Shrink the external surface. Inventory every internet-facing app, including partner portals, mobile backends and legacy services. Restrict cloud admin, SCM/Kudu and backup interfaces to trusted networks (FBI).
  • Harden identity. Apply least privilege to multi-tenant application permissions, and monitor credential additions and anomalous app-only activity (FBI).
  • Fix the basics. Remove hardcoded secrets, rotate keys, enforce authentication on every API route, and rate-limit sensitive endpoints (FBI; Datadog).
  • Detect behavior, not names. Talos advises treating tool signatures as leads and keeping detections for web shells, remote access tools, command-and-control and lateral movement (Talos).
  • Use decoys. Microsoft reports that honeypots, tarpits and fake wins helped pull agent traffic away from real assets in its testing (Microsoft).
  • Prepare to contain fast. Have a plan to revoke credentials and tokens, disable suspicious applications and isolate systems (Talos).

Limits of what we know

  • Scale is uncertain. Researchers found more than 1,600 exposed ARTEX consoles in early November, but the FBI says total deployments are likely much higher (FBI).
  • Numbers differ by source. Attempted sign-ins are not confirmed compromises, and the South Korea victim count is reported as about 67,000 by the FBI and at least 68,000 by TechCrunch.
  • Behavior varies. Datadog notes that the distributions it analyzed lack an opinionated cloud post-exploitation playbook, so activity depends on the model in use (Datadog).
  • Attribution is vendor assessment. Links to Storm-0940 or Storm-3358 are Microsoft's assessments (Microsoft; Microsoft).
  • Speculation: Talos expects future attack chains to include autonomous exfiltration and ransomware. That is a forecast, not an observed event (Talos).
  • We did not independently test these indicators. Validate them against your own telemetry.

Sources

Get the investigator briefing

Upcoming webinars (many with CPE), trends we're seeing in digital forensics and OSINT, and new articles like this one. About once a week.

Free. Unsubscribe anytime. We never share your email with sponsors.

Discussion

Comments are for members of The Hub. It's free to join.