Executive Summary
Between October 3 and mid-November 2026, Amazon's MadPot threat intelligence team captured records from autonomous AI agents running on ARTEX, an open-source AI-powered penetration-testing framework, targeting internet-exposed AI applications including LiteLLM, One API, Dify, Langflow, RAGFlow, and others. The agents exploited default credentials, unauthenticated endpoints, authentication-bypass headers, and known CVEs to harvest AI provider API keys at scale, then verified and pooled those keys to fund further reconnaissance. China-linked operators were assessed with moderate confidence based on Simplified Chinese prompts and China time-zone metadata. Hundreds of targets across AI platforms, CI/CD pipelines, and cryptocurrency exchanges were enumerated, with Cloudflare separately confirming 400 database dump attempts in a single hour from a compromised CI/CD token.
This campaign illustrates a structural shift in attacker economics: AI agents now automate vulnerability discovery, credential harvesting, and key validation end-to-end, dramatically compressing the time between exposure and exploitation. The FBI issued a FLASH alert confirming ARTEX use across multiple sectors globally since at least August 2026, underscoring that agentic AI offensive tooling has moved from proof-of-concept to mainstream criminal infrastructure.
Why This Matters Now
ARTEX and derivative frameworks are actively scanning thousands of organizations daily—over 50 AI pentesting framework families targeted Cloudflare customers in a single three-day window. With the original repository forked into commercial variants and community archives, takedown efforts have failed. Any organization running AI middleware, LLM gateways, or inference endpoints with default credentials or unpatched CVEs faces near-real-time automated exploitation, making immediate remediation and runtime network controls urgent priorities.
Attack Path Analysis
China-linked threat actors deployed ARTEX-based AI agents to autonomously scan internet-exposed AI gateway applications (LiteLLM, One API, Dify, RAGFlow, Langflow) for default credentials, unpatched CVEs, and misconfigured endpoints, extracting embedded API keys at scale. Stolen keys were immediately weaponized to fund further AI-powered reconnaissance, with exfiltrated credentials pooled across campaigns and used to access managed AI inference services, secret stores, and application identities, resulting in unauthorized AI model consumption, potential data exposure, and persistent access via CI/CD token abuse.
Kill Chain Progression
This analysis maps confirmed threat intelligence to the full cloud kill chain to show where defensive gaps would emerge as an attack progresses.
Initial Compromise
Description
ARTEX AI agents performed automated reconnaissance against internet-exposed AI application endpoints including LiteLLM, One API, Dify, Langflow, and RAGFlow. Entry vectors included exploitation of CVE-2025-3248 (Langflow unauthenticated RCE), CVE-2026-28797 (RAGFlow SSTI), One API authentication bypass via isCliRequest header, default or hard-coded administrator credentials (e.g., LiteLLM default master key), incomplete installations, and secrets exposed in client-side JavaScript, source maps, debug routes, environment files, and public repositories. Internet scanning services, certificate transparency logs, GitHub, npm, PyPI, container image layers, and Telegram channels were used to identify targets and leaked keys.
Related CVEs
CVE-2025-3248
CVSS 9.8A missing authentication vulnerability in Langflow allows a remote unauthenticated attacker to execute arbitrary code on the server via the /api/v1/validate/code endpoint, leading to full system compromise.
Affected Products:
Langflow Langflow – < 1.3.0
Exploit Status:
exploited in the wild
MITRE ATT&CK® Techniques
Valid Accounts
Unsecured Credentials
Unsecured Credentials: Credentials In Files
Exploit Public-Facing Application
Command and Scripting Interpreter
Acquire Infrastructure: Web Services
Data from Cloud Storage
Create Account
Adversary-in-the-Middle
Serverless Execution
Cloud Service Discovery
Cloud Service Dashboard
Reflective Code Loading
Obtain Capabilities: Tool
Gather Victim Org Information
Potential Compliance Exposure
Mapping incident impact across multiple compliance frameworks.
PCI DSS 4.0 – Authentication Management and Default Credential Prohibition
Control ID: Requirement 8.3.1 / 8.6.1
NYDFS 23 NYCRR 500 – Monitoring of Authorized Users and Password Management
Control ID: Section 500.14 / 500.7
DORA (Digital Operational Resilience Act) – ICT Security and Threat-Led Penetration Testing Controls
Control ID: Article 9 / Article 10
CISA Zero Trust Maturity Model 2.0 – Privileged Access and Credential Lifecycle Management
Control ID: Identity Pillar – Level 2 / 3: Least Privilege and Continuous Validation
NIS2 Directive – Policies on Risk Analysis, Access Control, and Supply Chain Security
Control ID: Article 21(2)(a) / 21(2)(e)
ISO/IEC 27001:2022 – Secure Authentication, Configuration Management, and Secure Development Lifecycle
Control ID: Annex A 8.5 / 8.9 / 8.25
Sector Implications
Industry-specific impact of the vulnerabilities, including operational, regulatory, and cloud security risks.
Information Technology/IT
AI agent campaigns targeting LiteLLM, Dify, and API gateways expose IT infrastructure to credential theft and cloud account takeover, with lateral movement and egress exfiltration risks mapped to NIST.800-53.SC-7 and ZTMM.Network controls.
Financial Services
ARTEX-driven AI credential harvesting directly targeted South Korean financial institutions, compromising 67,000 records; static CI/CD tokens and unpatched inference endpoints amplify PCI.4.0 compliance violations and account takeover exposure.
Internet
Internet-exposed AI app platforms including Open WebUI, LobeChat, and Ollama face automated agentic scanning, SSRF exploitation, and API key pooling attacks, undermining ZTMM.Applications and NIST.CSF.PR.AC-4 segmentation controls.
Computer/Network Security
Security tooling built on LLM gateways and AI penetration frameworks like ARTEX PRO is weaponized against defenders; compromised CI/CD tokens and missing egress policy enforcement violate ZTMM.CrossCutting.Governance and NIST.800-53.SC-18 requirements.
Sources
- Amazon Uncovers AI Agent Campaign Targeting LiteLLM, Dify, and Other Apps for API Keyshttps://thehackernews.com/2026/11/amazon-uncovers-ai-agent-campaign.htmlVerified
- Amazon Threat Intelligence Observes Threat Actors Running Cybercrime Campaigns with ARTEX AI Agentshttps://aws.amazon.com/blogs/security/amazon-threat-intelligence-observes-threat-actors-running-cybercrime-campaigns-with-artex-ai-agents/Verified
- Threat Actors Use Agentic AI Tooling to Find and Abuse Exposed Secretshttps://www.microsoft.com/en-us/security/blog/2026/11/05/threat-actors-use-agentic-ai-tooling-to-find-and-abuse-exposed-secrets/Verified
- Adapting to AI Penetration Testing Orchestrators - Cloudflare Bloghttps://blog.cloudflare.com/adapting-to-ai-penetration-testing-orchestrators/Verified
- FBI FLASH Alert - ARTEX AI Agent Abuse Across Sectorshttps://www.ic3.gov/CSA/2026/261113.pdfVerified
- NVD Entry - CVE-2025-3248 Langflow Missing Authentication RCEhttps://nvd.nist.gov/vuln/detail/CVE-2025-3248Verified
- CISA Known Exploited Vulnerabilities Cataloghttps://www.cisa.gov/known-exploited-vulnerabilities-catalogVerified
- ARTEX and Beyond - Cisco Talos Intelligence Bloghttps://blog.talosintelligence.com/artex-and-beyond/Verified
- When AI Infrastructure Becomes the Target: Securing Gateways and Control Pointshttps://www.microsoft.com/en-us/security/blog/2026/08/26/when-ai-infrastructure-becomes-target-securing-gateways-control-points/Verified
Frequently Asked Questions
Cloud Native Security Fabric Mitigations and ControlsCNSF
Based on the attack progression modeled above, these are the defensive controls that would constrain each stage.
Aviatrix Zero Trust CNSF is directly relevant to this campaign because ARTEX agents exploited internet-exposed AI application endpoints and then pivoted freely across cloud environments using stolen credentials and unrestricted network paths. Enforcing workload-level segmentation and identity-aware access boundaries would likely reduce the reachability of these exposed services and constrain the blast radius available to automated credential-harvesting agents.
Control: Cloud Native Security Fabric (CNSF)
Mitigation: Reducing direct internet reachability of AI gateway workloads through fabric-enforced perimeter segmentation would likely constrain the exposure surface available to automated ARTEX scanning agents, limiting which services could be directly probed for default credentials or unpatched vulnerabilities.
Control: Zero Trust Segmentation
Mitigation: Zero Trust segmentation enforced at the workload and network layer would likely reduce the ability of a compromised application instance to reach cloud metadata endpoints or internal credential stores, constraining the escalation path from application-layer access toward cloud control-plane permissions.
Control: East-West Traffic Security
Mitigation: East-west traffic enforcement across cloud environments would likely reduce the ability of compromised workloads to freely communicate with adjacent cloud services or account resources, constraining lateral pivot paths that relied on static CI/CD tokens and unfiltered cross-environment connectivity.
Control: Multicloud Visibility & Control
Mitigation: Multicloud visibility and control over outbound traffic destinations would likely surface anomalous high-frequency communications to unexpected external LLM infrastructure, potentially enabling policy enforcement that could constrain the ability of compromised workloads to maintain reliable orchestration callbacks to attacker-controlled endpoints.
Control: Egress Security & Policy Enforcement
Mitigation: Enforcing egress policies that restrict which external destinations AI application workloads are permitted to reach over HTTPS would likely reduce the ability of compromised services to transmit harvested credentials and secrets to attacker-controlled pooling infrastructure, constraining the exfiltration paths available to ARTEX agents.
Where prior segmentation, east-west enforcement, and egress controls would have constrained attacker reach, residual impact would likely be scoped to workloads and credentials already reachable within breached segments, reducing the breadth of unauthorized AI inference consumption and limiting credential pooling to a narrower set of harvested keys.
Impact at a Glance
Affected Business Functions
- AI Inference and Model Gateway Operations
- Cloud Credential and Secret Management
- CI/CD Pipeline Integrity
- Third-Party API and SaaS Integration Security
Estimated downtime: 7 days
Estimated loss: N/A
Mass harvesting of AI provider API keys (OpenAI, Anthropic, and similar inference providers) stored in exposed LLM gateways and proxy applications including LiteLLM, One API, New API, Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat, and Ollama. Cloud credentials including managed AI service identities and secret store contents were targeted. CI/CD tokens with broad account-level permissions were compromised enabling database dump attempts across customer environments. Hundreds of targets were enumerated across AI app platforms. Stolen keys were immediately weaponized to fund further attacker AI inference workloads, generating unauthorized compute spend for victims. Additionally, a secondary campaign involved deepfake-based identity bypass attempts at cryptocurrency exchanges, posing risk to customer identity verification data. Personal information on approximately 67,000 individuals was reported stolen in a related South Korean financial sector campaign attributed to the same ARTEX tooling.
Recommended Actions
Key Takeaways & Next Steps
- • Apply CNSF Egress Security and Cloud Firewall (ACF) controls with FQDN-based filtering to block unauthorized outbound connections from AI application workloads to unknown or attacker-controlled LLM endpoints, preventing exfiltration of harvested API keys and stopping agent callback communications to external orchestrators.
- • Deploy Zero Trust Segmentation and East-West Traffic Security to enforce least-privilege, identity-based policies between AI gateway services, model inference endpoints, and cloud management planes, ensuring that a single compromised AI app cannot be used to pivot into secret stores, CI/CD systems, or adjacent cloud accounts.
- • Enable Multicloud Visibility and Control with centralized traffic observability to baseline normal AI application behavior and generate alerts on anomalous patterns such as high-frequency API scanning (69+ req/min), mass database dump requests, sudden new workload deployments matching artex-* naming patterns, or unexpected cross-account token usage.
- • Enforce Egress Security policies requiring all AI inference traffic to exit only through approved, policy-controlled paths with application-layer inspection, preventing agents from using compromised keys to reach premium AI model services or cloud metadata endpoints via SSRF, and blocking outbound traffic to unauthorized destinations flagged as shadow AI or attacker infrastructure.
- • Leverage the Cloud Native Security Fabric (CNSF) with inline enforcement and real-time inspection to detect agentic AI tool behaviors including automated credential scanning, LLM-guided task orchestration traffic, and prompt injection patterns at the network layer, while enforcing runtime policy that restricts AI application containers and Kubernetes pods from accessing environment variable stores, source maps, or unauthenticated management endpoints.



