ARTEX: The AI Agent Framework Now Running Real Cybercrime Campaigns

AI Security Series

Between October 3 and November 5, 2026, Amazon's MadPot threat sensors — a globally distributed network of decoy systems configured to look like popular LLM inference APIs — captured something unusual. AI agents reaching out to what they believed were legitimate model endpoints handed over their own operational memory: prompt logs, task records, and campaign notes from active cybercrime operations. The attackers' tools had inadvertently documented themselves.

The framework behind those agents is ARTEX, an open-source agentic penetration testing tool that China-linked threat actors have been running against AI infrastructure at scale since at least August 2026. Amazon published its findings on November 13. Microsoft, Check Point, and Datadog had already confirmed in-the-wild ARTEX activity in October and early November. Google's Threat Intelligence Group named ARTEX as one of four active autonomous offensive frameworks in a report published November 16. This is not an emerging threat. It is a documented, multi-vendor-confirmed campaign pattern that is already past the honeypot stage.

How ARTEX Works

ARTEX is architecturally similar to legitimate agentic AI frameworks — which is precisely what makes it effective. An operator sets a target and a goal. A planner agent breaks the goal into discrete tasks and distributes them to worker agents. Each worker executes its task using real security tools, then sends results to an LLM that selects the next step. A verification layer checks outcomes. A compression agent maintains a persistent knowledge graph as shared memory across the session.

The agents divide work among planner, worker, verification, and compression roles, sharing a persistent knowledge graph as common memory. — Amazon Threat Intelligence

When ARTEX agents reached Amazon's decoy LLM endpoints expecting to receive model completions, they instead submitted their memory context to Amazon's threat analysts — who could read the campaign records directly. The ARTEX GitHub repository was taken down on October 8. Forks, archived copies, and a commercial version called ARTEX PRO remain available.

Campaign One: Stolen Keys Paying for More Theft

The most prominent ARTEX campaign was a large-scale theft of credentials to access AI models, organized as a self-sustaining operation in which stolen keys paid for the LLM inference powering further credential hunting. — Amazon Threat Intelligence

The loop is structurally elegant from an attacker's perspective: no need to pay for model access when the previous round of stolen keys covers the cost of the next round. Targets included LLM gateways — One API, New API, LiteLLM — and AI application builders including Dify, Langflow, Ollama, and Open WebUI. — Amazon Threat Intelligence

The exploitation playbook combined default admin credentials, incomplete installations, known CVEs (including CVE-2025-3248 in Langflow and CVE-2026-28797 in RAGFlow), SSRF to cloud metadata services, and secrets left in exposed configuration files and client-side JavaScript.

Keys were sourced through GitHub, package registries, container images, paste sites, and Telegram. Stolen credentials were graded L1 to L3, with keys exposed more than 30 days treated as low-value and deprioritized. — Amazon Threat Intelligence

The grading system indicates a mature operational workflow, not opportunistic scanning. The Microsoft confirmation adds the sharpest data point on response windows: in at least one case, Microsoft observed an AI agent move from scanning GitHub to exfiltrating data from Azure Key Vault in five minutes. — Microsoft Threat Intelligence That is not a window that most security operations centers can respond within.

Langflow: A Repeat Target Problem

Langflow appears throughout the ARTEX campaign records as a primary exploitation target — and the vulnerability history explains why. Industry reporting indicates Langflow saw twelve vulnerabilities exploited in the wild during 2026 alone, compared with only one known exploited Langflow flaw before that year, and total exploitation attempts across the platform's 2026 CVE portfolio now exceed 15,000. — CSA Research

The pattern is consistent across every Langflow CVE that has seen active exploitation: unauthenticated access to an endpoint leads directly to credential harvesting. CVE-2026-0768, carrying a CVSS score of 9.8, had been sitting in Langflow's codebase since before its public disclosure as a zero-day in January 2026, with exploitation starting in late August and not slowing down since. — CSO Online

Langflow's architectural exposure is structural, not incidental. Because it connects to model providers, internal APIs, credentials, file systems, and automation pipelines, a vulnerability in Langflow is a production infrastructure risk — not a developer tool issue. The latest recommended patch target is version 1.11.6, which addresses all currently known vulnerabilities. Any Langflow instance running in a development or staging environment that hasn't been updated this quarter should be treated as potentially compromised.

Campaign Two: Deepfake KYC Bypass

The second campaign Amazon documented is distinct in target and method but runs on the same ARTEX infrastructure. Agents used real-time deepfake video against live selfie checks and forged ID photos against document checks at two cryptocurrency exchanges. They reverse-engineered one vendor's liveness detection protocol, running a deepfake pipeline at 11.5 frames per second on a single consumer GPU. Forged documents included synthetic camera metadata and preserved original compression tables to make them appear as unmodified camera photos. — Amazon Threat Intelligence

The good news in the Amazon data: the best deepfake attempts were rejected by the vendor's server-side verdict in every recorded case. The liveness check held. The bad news is that this is a documented, automated, agentic KYC bypass pipeline running against live production systems — and it was built against cryptocurrency exchanges, not healthcare. The same capability applied to patient identity verification or provider credentialing systems is a materially different risk profile.

What This Means for Healthcare

Healthcare organizations may not be running cryptocurrency exchanges, but they are running the exact AI infrastructure ARTEX was designed to exploit.

AI Application Infrastructure as a Target Surface

Langflow, LiteLLM, Dify, and Open WebUI are increasingly present in healthcare development environments — used to build RAG pipelines over clinical documentation, prototype AI-assisted workflows, and evaluate model performance on internal datasets. Amazon's core defensive recommendation is to apply production-grade controls to AI applications, model gateways, and inference endpoints — the same standard applied to any internet-facing production system. — Amazon Threat Intelligence

That framing is worth internalizing: a Langflow instance in a clinical IT development environment is not a low-stakes developer toy. If it has credentials, cloud access, or connectivity to systems that touch PHI-adjacent infrastructure, it carries production-grade risk. Healthcare security teams should inventory every AI application builder and model gateway in their environment — including developer workstations and staging servers — and confirm current patch status against the Langflow CVE list. Any instance with network reachability and an unpatched critical CVE should be treated as a priority remediation item, not a backlog ticket.

AI API Key Hygiene

In three recent LLMjacking cases, unauthorized AI usage costs ran to nearly $1 million for one organization, $25,000 for a software architect, and $600,000 in AI credits for an AI testing organization due to a stolen API key. — Okta Threat Intelligence

The ARTEX grading system — L1 through L3 based on key age and validity — means active keys are the highest-value targets. Healthcare organizations using AI model APIs should have controls in place for: key rotation schedules, usage alerts for unexpected spend, IP restrictions on API tokens, and secrets scanning across code repositories and CI/CD pipelines.

The specific Cloudflare finding from the Amazon report is also worth noting for any organization using Cloudflare Workers in clinical automation workflows: static CI/CD tokens that never expire were the root cause in several confirmed ARTEX incidents, and attackers created Worker scripts with artex-* naming patterns. Hunting for those naming patterns is a low-cost, high-signal detection step.

Deepfake Identity Verification as a Forward Threat

The KYC bypass campaign failed against its current targets — but represents documented capability at scale. Healthcare has two identity verification surfaces worth watching in this context: patient identity verification in telehealth and digital health enrollment flows, and provider credentialing systems that rely on document and liveness checks. Neither of these is a current ARTEX target. Both are plausible future targets as this tooling matures and migrates beyond financial services. Healthcare security and identity teams should be tracking liveness detection vendor capabilities and evaluating whether current verification controls would hold against a consumer-GPU deepfake pipeline.

The Bigger Picture

ARTEX is one framework. Google GTIG named three others — CAIRN, CyberStrikeAI, and PentAGI — in its November 16 landscape report, alongside ZeroHollow (an inference service used for exploit development) and underground LLM resale markets (APIMart, JetNight) where stolen AI credentials are actively traded. This is an ecosystem, not an isolated tool.

The Amazon report's headline finding — that agents handed over their own memory to a decoy — is a useful reminder that agentic systems behave differently from traditional attack tools. They carry context. They make decisions. They produce records. That creates intelligence collection opportunities that didn't exist with static malware, and it also creates attribution complexity that traditional IOC-based detection wasn't designed to handle.

Amazon's conclusion is unambiguous: the most effective defenses remain the fundamentals — no default credentials, no secrets in code, no unauthenticated endpoints, no unpatched software. — Amazon Threat Intelligence The sophistication of the attacker tooling makes those fundamentals feel insufficient. They aren't. ARTEX's entire exploitation playbook runs on hygiene failures. Fix the hygiene and the agentic framework has nothing to work with.


This is an entry in the AI Security series. For related coverage, see When AI Writes the Code, Who Checks the Supply Chain? and The Bottleneck Just Moved to Security. Is Healthcare Ready?


Key Links