
FBI Warns ARTEX AI Hacking Tool Is Driving Global Intrusions
.webp)
Federal investigators have put a name to the automation behind a run of unusually fast intrusions. In a FLASH advisory published on November 13, 2026, the FBI described ARTEX, an AI hacking tool built as an autonomous penetration testing framework. A range of threat actors have run it against organizations worldwide since at least August. Some reached privileged access in under three hours, leaving almost no window to spot the activity, let alone contain it.
The 13-page advisory, designated FLASH-20261113-001, maps observed techniques to MITRE ATT&CK and lists indicators of compromise. It cites no CVEs. Attackers running the ARTEX AI hacking tool lean on known weaknesses and misconfigurations. That puts the defensive gap in inventory and permissions, not patch cycles.
What the Framework Actually Does
ARTEX runs as an open-source multi-agent system written in Go. An operator supplies an objective in plain language, and the framework turns it into machine commands for specialized agents. Those agents handle reconnaissance, vulnerability discovery, exploitation, privilege escalation and data theft with little further human input. ARTEX ships no model of its own, which makes the AI hacking tool dependent on an outside provider.
Operators have pointed the AI hacking tool at Claude, GPT, DeepSeek, GLM, Grok and Kimi, and ARTEX treats each as an interchangeable backend. Chinese-developed models dominate attacker-controlled deployments, led by the DeepSeek Flash family. The FBI cited cost, capability and absent guardrails as the draw. It also called provider safety controls no brake defenders can count on.
A developer using the handle Autumn-27 published it on GitHub in July 2026, then pulled the repository on October 8. That removal, prompted by abuse reports, barely dented availability. Community forks, archived copies and a commercial edition sold as ARTEX PRO keep the AI hacking tool in circulation. Censys counted more than 1,600 exposed ARTEX consoles in early November, and the FBI believes the true figure runs higher.
Inside the ConnectWise Supply Chain Compromise
Microsoft attributed October’s ConnectWise Cloud Backup compromise to Storm-0940, a China-state-aligned actor it has tracked since 2021. Microsoft assesses that the group used the ARTEX AI hacking tool to accelerate scanning, password spraying and OAuth application abuse. The actor stole application identities belonging to the backup service, formerly SkyKick. Signing in as the application itself let it reach customer Microsoft 365 tenants with no MFA prompt in the way.
The headline figure deserves care, because attempted access and confirmed access are not the same thing. Microsoft recorded sign-in attempts against more than 2,250 tenants at rates reaching 2,900 per minute, and most failed. Huntress confirmed unauthorized access in at least 162 tenants, and watched ConnectWise service principals grant themselves directory roles in close to 50. One of those principals claimed Global Administrator.
ConnectWise disabled seven application identities and took the service offline on October 17, four days before Microsoft published its findings. Credentials the actor planted during the campaign outlive that shutdown. The FBI told Cloud Backup customers to review every affected tenant by application ID back to October 1. It also asked them to preserve Entra, Purview and SIEM evidence first, because fast containment destroys the record.
Where the ARTEX AI Hacking Tool Finds Its Openings
Much of the framework’s effectiveness comes from aiming at systems nobody clearly owns. The FBI listed partner portals, employee-facing applications, legacy services and third-party integrations as the surfaces where the ARTEX AI hacking tool finds vulnerabilities most reliably. Those assets sit outside the main patch program, and agents enumerate them at a pace no human team would attempt.
The logged behavior reads as methodical rather than exotic. Agents ran sub-second checks against Azure credential permissions, SAS tokens and storage keys, then enumerated users and mailboxes through Microsoft Graph. Initial access came through unsigned SAML assertions, injection flaws, hardcoded admin passwords in JavaScript bundles and stolen VPN configurations. Escalation followed familiar Active Directory paths, including Kerberoasting and DCSync to lift the krbtgt key.
One discovery sequence walked more than 110,000 directory objects through PowerShell queries. The same run read LAPS passwords and inventoried endpoint protection to find hosts without a sensor. None of it needed a zero-day.
Other Campaigns Tied to the Framework
South Korea absorbed some of the heaviest activity on record. A Chinese-speaking actor paired the ARTEX AI hacking tool with ClaudeCode agents against Shinhan Bank, KB Kookmin Bank and Hana Bank. Those intrusions exposed personal information on roughly 67,000 people. Police have since arrested seven suspects linked to attacks on more than 90 organizations.
CrowdStrike separately tracked the espionage group VaporPanda using the ARTEX AI hacking tool against more than 200 organizations in 14 countries. At least nine of them fell. Attribution across these campaigns remains uneven, which matters for anyone building defenses around one named adversary. The FBI named no state sponsor of its own, treating the framework as a widely circulated capability rather than one group’s signature.
What Still Worked
Nothing in the advisory casts the automation as unstoppable. Endpoint protection blocked Pass-the-Hash attempts wherever a functioning sensor sat in place. That detail matters, because the ARTEX AI hacking tool spent part of its reconnaissance effort mapping which hosts lacked exactly that coverage.
The recommended response starts with something unglamorous, a full inventory of every internet-facing application. Management planes belong off the public internet, covering cloud administration consoles, Kudu and SCM endpoints, hypervisor interfaces and backup systems. Least privilege matters most for multi-tenant application permissions, where one over-scoped app can open hundreds of downstream environments. Rate limits blunt the agents’ sequential enumeration, and hunting signals include unexplained LLM API traffic and sudden AI billing spikes.
The Window Has Narrowed
What the advisory describes amounts to a change in tempo rather than a change in technique. The vulnerabilities the ARTEX AI hacking tool exploits were already present, already documented and already fixable. What has collapsed is the schedule, which most organizations built around attackers working at human speed.
Three hours from first contact to privileged access leaves no room for a weekly review cycle. Every item on the FBI’s mitigation list was sound advice before ARTEX existed, and none of it requires new tooling. Organizations that know precisely what they expose, and who holds permissions over it, keep the only advantage that still scales.
Subscribe to receive the latest blog posts to your inbox every week.