FBI Warns ARTEX AI Agents Compress Cyberattacks Into Hours
The FBI says ARTEX, an open-source AI agent framework, has driven intrusions reaching privileged access in under three hours across 11 sectors since August.

Why it matters
- Some ARTEX-driven intrusions reached privileged access in under three hours, the FBI said in a November 13 FLASH alert.
- The FBI puts exposed ARTEX consoles at more than 1,600 as of early November; Censys counted 1,465 on November 4.
- CrowdStrike tracked Vapor Panda using ARTEX against more than 200 organizations in at least 14 countries from October 17.
- The ConnectWise supply-chain compromise enabled access to more than 2,250 downstream tenants, the FBI said.
- Amazon's MadPot sensors caught ARTEX agents between October 3 and November 5 stealing AI model credentials.
The FBI says attackers using ARTEX, an open-source AI agent framework, have compressed intrusions from initial targeting to privileged access into under three hours. In a FLASH alert released Friday, the bureau said threat actors have turned the penetration-testing tool into a way to "automate multiple phases of cyber intrusions at machine speed, from reconnaissance through data exfiltration."
The FBI Cyber Branch announced the warning publicly on November 13, writing on X that "a variety of malicious cyber actors have used ARTEX to conduct campaigns" against organizations worldwide. The stakes are straightforward: the same agentic AI capability that vendors sell for legitimate automation now runs attack chains that defenders struggle to answer in time.
What is ARTEX and why did it survive its own takedown?
ARTEX is written in Go. It takes goals written in plain language and converts them into commands executed by specialized agents, with little human involvement once the objective is set. The framework has no model of its own. It plugs into outside ones, from Anthropic's Claude and OpenAI's GPT to open-source and Chinese-origin models such as DeepSeek, GLM and Kimi.
In the attacker-controlled instances the FBI analyzed, Chinese models were the most common. DeepSeek's Flash family dominated, the alert says, "due to their balance of capabilities, relatively low cost, and lack of observable guardrails when framed for penetration testing."
Internet-scanning firm Censys, whose research the bureau cites, laid out the mechanics earlier this month. A planner agent breaks the goal into attack "intents." Worker agents claim those intents, run tools such as nmap, write the results back, and the loop repeats until the goal is met.
The framework first appeared on GitHub in July. Its developer said the tool was meant for authorized testing within defined scopes and made the repository unavailable on October 8, after reports of attacker use. That hasn't contained it. Community forks, localized variations, archived copies and a commercial version called ARTEX PRO keep it in circulation.
Censys counted 1,465 hosts exposing an ARTEX console on November 4 and noted the console doesn't need to face the internet at all, so total deployments are likely higher. The FBI's alert puts the exposed count at more than 1,600 as of early November.
Which campaigns has ARTEX powered?
The alert pulls together campaigns that security vendors have documented since early October. Microsoft attributed the ConnectWise Cloud Backup supply-chain compromise to Storm-0940, a group it describes as China-state-aligned, and said the group used ARTEX to speed up scanning, password spraying and abuse of OAuth applications. The FBI says that compromise enabled access to more than 2,250 downstream tenants, including tenants that no longer used the backup service.
A separate, financially motivated campaign against South Korean financial institutions exfiltrated personal information on about 67,000 people, the bureau said.
CrowdStrike has tracked some of the largest operations. The company said a group it calls Vapor Panda ran ARTEX primarily with DeepSeek against more than 200 organizations in at least 14 countries starting October 17, and it confirmed post-exploitation activity at nine of them. In a single intrusion, the group's planning agents consumed "on the order of 1.4 billion input tokens," CrowdStrike said.
Amazon added its own findings Friday. Its MadPot decoy sensors, configured to look like popular LLM inference APIs, caught ARTEX agents sending along their own campaign records between October 3 and November 5. The most prominent campaign Amazon described was large-scale theft of credentials for AI models, "organized as a self-sustaining operation in which stolen keys paid for the LLM inference powering further credential hunting."
The speed is what worries the bureau most. A variety of malicious actors have used the framework since at least August against organizations in financial services, government, technology, defense, semiconductors, energy, healthcare, manufacturing, telecommunications and law. ARTEX traffic is noisy enough to set off alarms, the alert notes, but the speed of the attacks leaves "limited time for coordinated defender response before objective completion."
AI is accelerating flaw discovery as well. VulnCheck says AI-driven bug hunting is behind most of the more than 100,000 CVEs published so far this year.
Where does the FBI want defenders to start?
The FBI wants companies to look hardest at the systems they watch least. It singled out partner and supplier portals, mobile backends, sales-support systems, legacy services and third-party integrations as the places the tool excels at breaking into.
Most of the entry points in the FBI's tactics table are familiar weaknesses, now worked through faster:
- Unsigned SAML assertions
- XXE and SSRF flaws
- Hard-coded administrator passwords in single-page-app JavaScript bundles
- Web APIs with no authentication
- Weak admin credentials and stolen VPN configurations
The post-access activity the bureau lists includes enumerating more than 110,000 Active Directory objects and taking inventory of which Windows machines lacked endpoint sensor coverage.
The bureau's first recommendation is an inventory of every internet-facing application, including the ones built by individual business units. Its further guidance:
- Restrict cloud administration, network device management, Kudu/SCM publishing endpoints, hypervisors and backup platforms to trusted networks or dedicated admin paths
- Apply least privilege to multi-tenant app permissions and consent grants
- Monitor credential additions to applications
- Set rate limits on high-impact business APIs
- Continuously scan code repositories, container images and CI/CD pipelines for exposed secrets
- On partner and employee portals, alert on sequential lookups and unusually high numbers of records pulled in a single session
What indicators should teams watch?
The indicator list runs to 24 IP addresses and one domain, an LLM API reseller tied to the South Korea campaign. The FBI cautions that cloud addresses get reassigned and should be vetted before anyone blocks them.
Behavioral indicators include high-frequency sequential requests against login and search endpoints, web shells in unexpected application paths, and unexpected outbound calls to commercial model APIs alongside other suspicious activity. Organizations reporting incidents are asked to include LLM API billing anomalies and any evidence of self-hosted model gateways or reseller software.
For ConnectWise customers, the alert is specific. The FBI says to review every Microsoft 365 tenant where the affected applications were ever installed or authorized, searching by application ID rather than display name, and to keep those connections disabled until ConnectWise provides formal restoration instructions.
With ARTEX forks and a paid PRO version still circulating after the original repository went dark, the bureau's message is that defenders can no longer assume human-paced attacks — the window for response is now measured in hours, not days.
Source: webpronews.com
More from Elena Vasquez
Show full bio
Market editor covering technology and product news in artificial intelligence and machine learning at AI In Context.
591 articles
Related articles
- CrowdStrike: AI Hacking Tool Let One Attacker Breach Korean Banks
- One Israeli Startup Sits Behind a String of Rogue AI Disclosures
- Five Agent Vulnerabilities in Five Months Expose Structural Flaw in MCP
- One Hacker, a Few Thousand Dollars, and an AI Army That Breached John Deere
- Anthropic's Red Team Built an AI Worm That Hid Across 15 TVs