Hackers Are Using the ARTEX AI Framework: From a Leaked Key to Azure Secrets in Five Minutes
ARTEX is an open-source AI framework for simulated intrusion testing, published on GitHub in July 2026, and real attackers are already using it. A planner agent hands tasks to other agents, and Microsoft has seen them find leaked keys in a public repository and extract data from Azure Key Vault, the secrets store, within five minutes. That is why a leaked key should be revoked immediately, not on a schedule.
AI-processed from X: @DFIR_Radar; edited by Hamidun News
Imagine a developer accidentally commits an access key for a cloud application to a public GitHub repository. Previously, an attacker had to find it, figure out what it unlocked, and manually work through the infrastructure. Now a team of AI agents does this work.
Microsoft recorded a case where only five minutes passed between finding the key and extracting data from Azure Key Vault, a cloud store for passwords and secrets. The tool behind this is ARTEX, an open-source penetration testing framework that a developer using the pseudonym Autumn-27 published on GitHub in July 2026. Reports from Microsoft, CrowdStrike, Cisco Talos, Datadog, and Dosu show that a tool built to test defenses is already being used in real intrusions by China-linked groups and financially motivated criminals.
How ARTEX Works
The design will be familiar to anyone who has built multi-agent systems. An operator sets a target and the desired outcome. A planner agent breaks the task into steps, executor agents run security tools, and all findings feed into a shared graph. The framework works with different language models and needs almost no human involvement once it starts.
ARTEX has no built-in C2 server, the channel through which an attacker controls infected machines. But operators can add anything they want through MCP extensions, the standard for connecting external tools to AI agents. Talos has seen ARTEX deployed alongside VShell, GoPhish, and AdaptixC2. Similar projects exist too: Hephaestus is controlled through Telegram, PentAGI has a web interface (Talos found more than 700 instances reachable from the internet), and HexStrike AI connects to MCP clients such as Claude Code.
What Has Already Been Done With It
CrowdStrike links the espionage campaign that began on October 17, 2026, to the Chinese group VAPOR PANDA. The campaign targeted more than 200 organizations in at least 14 countries. Breaches are confirmed at nine of them, and at least 12 more were subjected to mass password guessing. Microsoft attributes attempts to sign in to more than 2250 Microsoft 365 tenants (a tenant is a separate account environment belonging to a customer company) to the group Storm-0940, after it compromised ConnectWise application credentials. Storm-3358, described as a financially motivated group from Cambodia, attacked financial organizations and then deployed WISP, VShell, and AdaptixC2.
There is also an ambiguous case. A group believed to be Chinese-speaking exfiltrated data from South Korean financial companies. Talos found traces there linked to ARTEX: directory listings and references to prompt history. The executable and source code themselves were absent, so this points to an ARTEX environment but does not prove the entire framework was used.
The speed is striking. Microsoft saw agents search public GitHub repositories for exposed application credentials and AI API keys. Datadog recorded anonymous requests to S3 in AWS, use of compromised long-lived keys, searches of CloudWatch logs for financial data and secrets, and registration of accounts in public Cognito pools to harvest tokens. Dosu described an attacker who moved from an authorization vulnerability to CI, a database, and the cloud, sending more than 80,000 requests in 90 minutes.
Caveats and What Defenders Should Do
The reports ask that the framework not be credited with more than it does. Post-breach actions depend on the specific group and are not built into ARTEX by default. CrowdStrike describes how VAPOR PANDA forged an administrator token from historical authentication data, read mail, and reached the internal network through VPN data it had found, and used web shells and the BLOODALCHEMY backdoor. Talos notes that ARTEX has already demonstrated evasion of EDR (endpoint detection and response) software.
Talos identified more than 2500 C2 addresses linked to ARTEX and examined which tools appear on the same servers. The most common are ScopeWeaver, AntiFox, CyberStrikeAI, PentAGI, and ScopeSentry, followed by VShell, Nessus, Interactsh, Acunetix, Fscan, NPS, AdaptixC2, Mimikatz, Cobalt Strike, Gophish, and Chisel. Co-occurrence does not prove that ARTEX invoked each of them, or that a single actor stood behind all of them.
The practical recommendations in the report come down to hunting for speed and identity abuse:
- correlate authentication logs, cloud API logs, and endpoint logs, looking for request bursts, unfamiliar applications, unusual token use, and changes to permissions and roles;
- check Microsoft Graph activity performed with application credentials, Key Vault access after secrets have been exposed, and registrations in Cognito;
- treat tool signatures and user-agent strings as leads, not proof, since they change quickly;
- keep independent telemetry in case EDR is disabled;
- rehearse mass revocation of credentials and tokens in advance.
Indicators of compromise are published in Talos's GitHub repository. The main lesson for developers: a key leaked to a public repository is now found and used not in days but in minutes, so secret scanning and short-lived tokens are no longer just good practice.
- Nov 21, 2026Microsoft outage: nearly 83,000 Outlook complaints on Downdetector, Teams and Copilot down too
- Nov 21, 2026MLPerf Training Adds Nvidia Rubin and TPU Ironwood in First Appearance: 153 Systems
- Nov 21, 2026Microsoft Turns Copilot Into an "Operating System for Work": Word and Excel Become Agent Plugins
- Nov 21, 2026Google to split Gemini chat history from model training, so you can opt out without losing your history
Key Questions Answered
How ARTEX Works?
The design will be familiar to anyone who has built multi-agent systems. An operator sets a target and the desired outcome. A planner agent breaks the task into steps, executor agents run security tools, and all findings feed into a shared graph. The framework works with different language models and needs almost no human involvement once it starts. ARTEX has no built-in C2 server, the channel through which an attacker controls infected machines.
What Has Already Been Done With It?
CrowdStrike links the espionage campaign that began on October 17, 2026, to the Chinese group VAPOR PANDA. The campaign targeted more than 200 organizations in at least 14 countries. Breaches are confirmed at nine of them, and at least 12 more were subjected to mass password guessing.
What are the key details on Caveats and What Defenders Should Do?
The reports ask that the framework not be credited with more than it does. Post-breach actions depend on the specific group and are not built into ARTEX by default. CrowdStrike describes how VAPOR PANDA forged an administrator token from historical authentication data, read mail, and reached the internal network through VPN data it had found, and used web shells and the BLOODALCHEMY backdoor. Talos notes that ARTEX has already demonstrated evasion of EDR (endpoint detection and response) software.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.