Autonomous AI Agent Campaigns Target Model Gateways and Tools Like LiteLLM, Dify, and Langflow (CVE-2025-3248, CVE-2026-28797)

Publication date: November 16, 2026
Category: Threat Intelligence / AI Cyberattacks

Introduction

Threat intelligence teams at Amazon and Cloudflare have uncovered a sophisticated global campaign in which autonomous malicious artificial intelligence agents systematically scan and target internet-exposed AI applications. Utilizing autonomous penetration testing frameworks such as the open-source tool ARTEX, operators seek to steal model provider API keys stored within gateways, proxies, and AI application development frameworks like LiteLLM, One API, Dify, Langflow, and RAGFlow. Such threats demonstrate how agent-based automation exponentially increases the speed and persistence of offensive operations in modern digital infrastructure.

What is the ARTEX Framework and the Exploited Vulnerabilities? (General Analysis)

ARTEX is an AI-agent-based penetration testing framework that operates via a task-split planning and execution architecture. A human operator defines a high-level goal, which a planner agent breaks down into discrete tasks assigned to worker agents interacting with security tools.

In these campaigns, agents process full execution histories and drive their decision-making using Large Language Models (LLMs), automating the search for misconfigurations, default credentials (such as LiteLLM’s default master master key), and known vulnerabilities within AI infrastructure components.

The analyzed vectors highlight two critical vulnerabilities documented across orchestration platforms:

  • CVE-2025-3248 (Langflow): Missing authentication vulnerability in versions prior to 1.3.0, allowing a remote unauthenticated attacker to execute arbitrary code. This issue is part of CISA’s Known Exploited Vulnerabilities (KEV) catalog due to active exploitation in ransomware campaigns (Reasoned CVSS v3.1 estimation: 9.8 Critical / CWE-306).
  • CVE-2026-28797 (RAGFlow): Server-Side Template Injection (SSTI) vulnerability in versions 0.24.0 and earlier enabling any authenticated user to execute operating system commands on the server (Reasoned CVSS v3.1 estimation: 8.8 High / CWE-94).

How Does It Work? (Technical Analysis)

The attack mechanism relies on automated cycles of reconnaissance, validation, and exfiltration executed by autonomous agents:

  • Decoy Infrastructure (MadPot): Through decoy server networks emulating popular LLM services, agents were observed sending payloads carrying operator instructions, progress logs, and detailed findings regarding exposed systems.
  • Exploitation and Credential Harvesting: Agents scan internet search engines, certificate transparency logs, GitHub repositories, npm and PyPI packages, and messaging channels to gather exposed API keys. They prioritize multi-provider gateways (such as One API, New API, Veloera, and LiteLLM) because a single compromised administrative panel grants access to numerous AI inference accounts.
  • CI/CD Token Abuse and Persistence: In incidents reviewed by Cloudflare, agents compromised static, account-level API tokens (devoid of expiration dates or IP restrictions) originating from third-party cloud provider integrations. Subsequently, they deployed malicious worker scripts under structured naming conventions (artex-*, artex-pgrecon-*, artex-pgw-*, artex-do-*) executing high-concurrency requests exceeding 69 requests per minute for mapping and session validation.

Affected Systems / Environments

Threat actors have targeted multiple LLM gateway platforms, AI application builders, and auxiliary services:

  • LLM Gateways and Proxies: One API, New API, Veloera, Sub2API, and LiteLLM.
  • AI Application Builders and Frameworks: Dify, FastGPT, Langflow (versions < 1.3.0), RAGFlow (versions <= 0.24.0), Open WebUI, LobeChat, and Ollama.
  • Corporate cloud environments and CI/CD pipelines utilizing static tokens without rate limits, geographic, or network restrictions.
CVE Category (CWE) Impact CVSS (Estimated) Vector (Summary)
CVE-2025-3248 CWE-306: Missing Authentication Remote Code Execution 9.8 (Critical) CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H
CVE-2026-28797 CWE-94: Code Injection (SSTI) OS Command Execution 8.8 (High) CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:H/I:H/A:H

Mitigation and Detection

Remediation

  • Upgrade Langflow to version 1.3.0 or later to mitigate CVE-2025-3248.
  • Upgrade RAGFlow to version 0.25.0 or later to ensure sandbox template rendering isolation and mitigate CVE-2026-28797.
  • Modify default administrative credentials (such as LiteLLM’s default master key) and prevent leaving unfinished or unauthenticated installations exposed.
  • Implement strict expiration dates and source client IP restrictions for all Cloudflare API tokens and cloud platform credentials.
  • Immediately rotate any AI provider keys exposed in public repositories or CI/CD logs.

Detection

  • Monitor log activity for suspicious Cloudflare Worker scripts matching the artex-* prefix.
  • Audit anomalous traffic patterns or recurring mass requests (such as database mapping sweeps or bulk session validations).
  • Implement Cisco Talos Snort detection rules to identify ARTEX agent framework activity.

“None of the techniques in these campaigns were new. What AI agents change is the speed and persistence of the attacker against misconfigurations.”

Wrapping Up

The autonomous AI agent campaign analyzed by Amazon and Cloudflare highlights a significant tactical evolution in the threat landscape: the deployment of agent-based penetration testing frameworks (such as ARTEX) to automate the discovery of exposed secrets and vulnerabilities across AI platforms. Insufficient hardening in LLM gateways and the misuse of static CI/CD tokens enable rapid intrusions threatening global corporate security. Adopting stringent token lifecycle controls and prompt software patching remains essential to mitigate this risk.

References