Autonomous Frameworks — Automated Offensive Campaigns with ARTEX: AI Agents at Scale
Publication date: November 16, 2026
Category: AI attacks (LLM/LocalAI)
Introduction
Amazon’s threat intelligence team has published critical findings regarding offensive campaigns executed by cybercriminals using ARTEX, an open-source agentic AI penetration testing framework. Between October 3 and November 5, Amazon’s MadPot decoy sensor network intercepted requests sent by these AI agents containing detailed logs of their operations. Although the threat actors automated credential harvesting, vulnerability exploitation across AI platforms, and biometric bypasses, analyzing these operational logs has shed light on the exact vectors and infrastructure used to scale malicious operations without constant human intervention.
What is ARTEX? (General Analysis)
ARTEX represents a significant milestone in the automation of cybercrime, evolving from traditional scripts to cooperative AI agent systems. The framework operates through a division of specialized roles: planner, worker, verification, and compression agents, all interconnected via a persistent knowledge graph. This architecture allows multiple instances to share discoveries and seamlessly resume tasks where others left off.
Within its operational playbook, the agents incorporated instructions to exploit configuration weaknesses and known vulnerabilities, notably:
- CVE-2025-3248 in Langflow (Severity estimation: High, CWE-79 or similar injection flaw).
- CVE-2026-28797 in RAGFlow (Severity estimation: High, CWE-20 improper input validation).
- Authentication bypasses in gateways like One API by abusing the
isCliRequestHTTP header. - Exploitation of Server-Side Request Forgery (SSRF) to access cloud provider metadata services.
Technical Note: CVSS scores and vectors for the CVEs mentioned in this analysis are considered reasoned estimations based on the architectural nature of the affected components (RAG platforms and AI pipelines), as official NVD records were not yet published at the time of writing.
How Does It Work? (Technical Analysis)
ARTEX stands out due to its operational autonomy and relentless persistence, structured across coordinated phases:
- Initial Reconnaissance and Automated Workflow: Operators define high-level missions (e.g., systematically scanning specific country lists, favoring enterprise production sites over personal projects) and delegate execution to the agents. These agents comb public code repositories, certificate transparency logs, npm/PyPI packages, and container layers for exposed API keys.
- Persistence, Escalation, and Credential Validation: Stolen keys are fed directly into the framework’s own model gateway to fund and empower agents hunting for additional credentials. Agents validate cloud identities, enumerate storage, target managed AI services, and grade assets based on potential impact (L1: lead, L2: working API call, L3: measurable impact like readable data or spendable balance).
- Biometric Bypass and KYC Fraud: In a secondary campaign targeting cryptocurrency exchanges, agents utilized real-time deepfakes running on a single consumer GPU (reaching 11.5 frames per second) to defeat live liveness checks, alongside forged identity documents featuring manipulated camera metadata to mimic untouched photos.
Affected Systems / Environments
The campaigns primarily targeted AI development platforms, API gateways, and environments prone to secret exposure.
| CVE / Component | Category (CWE) | Impact | CVSS (Estimated) | Vector (Summary) |
|---|---|---|---|---|
| Langflow (CVE-2025-3248) | CWE-79 (Estimated) | Command Execution / Injection | 8.5 (High) | CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:L |
| RAGFlow (CVE-2026-28797) | CWE-20 (Estimated) | Input Manipulation / Unauthorized Access | 8.1 (High) | CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:H/I:H/A:N |
| One API / New API / Veloera | CWE-287 (Authentication) | Header-based Auth Bypass | 7.5 (High) | CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:N/A:N |
| LiteLLM & Similar | CWE-798 (Default Credentials) | Default Master Key Exposure | 9.8 (Critical) | CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H |
Impacted Environments: Gateways including One API, New API, Veloera, Sub2API, and LiteLLM; development platforms such as Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat, and Ollama; as well as cloud storage and cryptocurrency infrastructure.
Mitigation and Detection
Remediation
- Removal of Default Credentials: Immediately rotate master keys and default passwords across LiteLLM and other inference gateways.
- Patch Management: Apply the latest security updates to patch known vulnerabilities in frameworks such as Langflow and RAGFlow.
- Secret Hygiene: Deploy automated scanning tools across code repositories and CI/CD pipelines to prevent accidental leaks of API keys and environment files (
.env). - AI Endpoint Hardening: Treat administrative LLM interfaces and inference endpoints with the same rigorous security standards applied to core production databases.
Detection
Defensive teams should actively monitor anomalies in inference resource consumption and unusual API call patterns.
“None of the techniques in these campaigns were new. What AI agents change is the speed and persistence of the attacker where basic security hygiene was missing.”
Suggested Detection Rule (HTTP header anomaly indicators):
title: Detection of isCliRequest Header Abuse in API Gateways
status: experimental
description: Identifies suspicious HTTP requests attempting to bypass authentication via CLI header manipulation.
logsource:
category: webserver
product: nginx
detection:
selection:
cs-method: "POST"
httpRequestHeader: "*isCliRequest*"
condition: selection
falsepositives:
- Legitimate command-line administration tools.
level: highWrapping Up
The utilization of the ARTEX framework by cybercrime groups marks a pivotal turning point in cybersecurity: the transition toward attacks orchestrated by autonomous AI agents capable of performing reconnaissance, credential validation, and biometric bypass in parallel. Far from relying on complex zero-day exploits, these campaigns capitalize on fundamental configuration flaws and poor operational hygiene. Effective defense requires hardening AI endpoints with the same rigor applied to critical corporate infrastructure.
References
- Security Affairs. (2026, November 16). ARTEX: How Cybercriminals Use AI Agents to Scale Their Attacks. https://securityaffairs.com/?p=203188