ARTEX AI
An open-source penetration testing assistance tool built on a planner-worker multi-agent architecture. It relies on external large language models and can automate the execution of network attack tasks.
Timeline (last 90 days)
On November 13, 2026, the FBI released a 13-page Flash warning stating that since at least August 2026, multiple attackers have used ARTEX to attack industries including banking, government, defense, healthcare, and telecommunications.
ARTEX AI was originally released on the code hosting platform by developer Autumn27 in July 2026, positioned as a penetration testing assistance tool.
ARTEX adopts a Planner-Worker multi-agent architecture: the Planner decomposes the target into subtasks, the Worker invokes conventional security scanning tools to execute and record, and the observe-decide-act loop runs repeatedly until the objective is achieved.
ARTEX itself has no built-in intelligence and relies on external large language models; the FBI listed six models that can be connected, including Claude, GPT, DeepSeek, and Grok, with DeepSeek being the most used by attackers.
Google's threat team observed that ARTEX is "much noisier" than a competent human tester, with a single attempt able to leave 240 log records, and one agent once sent emails to more than 1,300 employees.
In early November 2026, Census found 1,465 ARTEX control panels accessible from the public internet, while the FBI's figure was more than 1,600; about 650 appeared to have been mass-generated from the same template by two hosting companies.
All Facts (6)
ARTEX AI was originally released on the code hosting platform by developer Autumn27 in July 2026, positioned as a penetration testing assistance tool.
50%UnverifiedARTEX adopts a Planner-Worker multi-agent architecture: the Planner decomposes the target into subtasks, the Worker invokes conventional security scanning tools to execute and record, and the observe-decide-act loop runs repeatedly until the objective is achieved.
50%UnverifiedARTEX itself has no built-in intelligence and relies on external large language models; the FBI listed six models that can be connected, including Claude, GPT, DeepSeek, and Grok, with DeepSeek being the most used by attackers.
50%UnverifiedGoogle's threat team observed that ARTEX is "much noisier" than a competent human tester, with a single attempt able to leave 240 log records, and one agent once sent emails to more than 1,300 employees.
50%UnverifiedIn early November 2026, Census found 1,465 ARTEX control panels accessible from the public internet, while the FBI's figure was more than 1,600; about 650 appeared to have been mass-generated from the same template by two hosting companies.
50%UnverifiedOn November 13, 2026, the FBI released a 13-page Flash warning stating that since at least August 2026, multiple attackers have used ARTEX to attack industries including banking, government, defense, healthcare, and telecommunications.
50%