Published: 2026-11-18
Categories: Agentic AI Threats
Self-Funding Agents: ARTEX Campaigns Harvesting AI-Gateway Keys
Key Takeaways
Amazon Threat Intelligence reported on November 13, 2026, that its MadPot decoy network captured the operating records of AI agents running credential-theft and fraud campaigns, most likely on the open-source ARTEX framework, between October 3 and November 5, 2026 [1]. The decoys posed as LLM inference APIs. The attacking agents used them as model backends and sent full task context, tool histories, and campaign findings with their requests [1]. Five implications stand out for defenders.
- The most prominent campaign funded itself. According to Amazon, the agents put stolen AI provider keys into the framework’s own model gateway and used them to pay for the inference that drove the search for more keys [1].
- AI gateways and app builders were the main targets. The playbook prioritized applications that hold keys for several providers, including LiteLLM, One API, New API, Dify, Langflow, RAGFlow, Open WebUI, and Ollama [1][2].
- Amazon reports that none of the techniques were new. The agents relied on well-known weaknesses such as default credentials, unfinished installs, and known CVEs (summarized in the table below). In Amazon’s words, what agents change is the speed and persistence of the attacker [1].
- The reported outcomes are unverified. Amazon could not observe the activity directly or confirm the results the agents recorded. Attribution to separate Chinese-nexus groups is at moderate confidence [1].
- The decoys worked because the attackers relied on third-party inference. When an agent sends its model calls to an endpoint the defender controls, that endpoint receives the agent’s working memory, and Amazon’s results suggest that inference-shaped decoys are a promising intelligence source [1].
The sections that follow place these findings in the context of the broader ARTEX activity reported this autumn, examine why AI gateways concentrate risk, and set out the controls that address the specific conditions in the attackers’ playbook.
Background
ARTEX is an open-source, multi-agent penetration testing framework that turns a high-level operator goal into tasks carried out by LLM-driven agents [3][6][10]. It does not include a model of its own. Instead, it calls external LLMs, including Anthropic Claude, OpenAI GPT, DeepSeek, GLM, and Kimi, and the FBI reports that Chinese models, especially DeepSeek Flash-family models, are the most common in attacker-controlled instances [3][10]. The developer withdrew the original GitHub repository on October 8, 2026, after abuse reports. Forks, archived copies, and a commercial edition called ARTEX PRO remain available [2][3][10]. Wiz reported that the code had more than 12,000 forks by the time it was made closed-source, so removing the original did little to limit access [6][11].
The Amazon disclosure arrived during a period of heavy pressure on AI credentials. In October 2026, Wiz Research documented a series of incidents at AI developer platforms, which it called “Black October.” In these incidents, attackers targeted stored environment variables, especially those granting access to AI models, and stolen keys appeared in LLM resale operations [4][15]. CSA’s follow-up analysis cited Railway’s estimate that keys stolen from its customers drove about $740,000 in unauthorized AI usage [7]. On November 5, Microsoft reported that threat actors were using ARTEX to search public GitHub repositories for exposed secrets, including AI inference keys [2][12]. In one case Microsoft described, an agent went from a GitHub search to data exfiltration from Azure Key Vault in about five minutes [5][12]. Then, on November 13, the FBI issued a FLASH alert stating that ARTEX has been used since at least August 2026 and that researchers had identified more than 1,600 publicly exposed ARTEX consoles as of early November [3][10].
Against this backdrop, Amazon’s report adds visibility into the attackers’ own planning. MadPot is a globally distributed network of decoy sensors, and for this work it presented LLM-compatible chat-completions interfaces on a range of ports [1]. Agentic frameworks keep conversation history and resend it with each model call. A single request captured by a decoy therefore often contained the planner’s instructions, the workers’ tool outputs, the shared knowledge-graph state, and the agents’ own summaries of what they had found [1]. Amazon states only that the frameworks sent prompts to these interfaces as part of their offensive workflows; its report does not explain how the decoys came to be selected as backends [1]. In practice, the agents sent their campaign notes to a sensor they treated as a model provider. Amazon also reports that it notified potentially affected AWS customers as appropriate [1].
Security Analysis
The captured records reveal four things worth examining in turn: the economics that sustained the largest body of activity, the reasons AI gateways sat at the top of the target list, the speed and scoping of discovery, and a parallel fraud campaign that shows what cheap inference can support. A discussion of evidence limits and the defensive opportunity follows.
The Self-Funding Loop
For defenders, the most consequential finding is economic. Amazon writes that stolen keys went into the framework’s own model gateway and powered the agents looking for more keys, which made the operation self-sustaining [1]. In an agentic campaign, the largest variable cost is inference, because every planning step, tool-output review, and verification pass requires a model call. Harvested keys cover that cost, so the operator’s marginal inference cost likely approaches zero, and each newly validated key adds capacity to the operation. The victims carry the cost through charges billed to their provider accounts. CSA’s loss-allocation analysis documents one case in which stolen keys drove an estimated $740,000 in unauthorized AI usage, and shows that such charges can reach six figures before they are contained [7]. CSA’s earlier work on LLMjacking had already described stolen compute being repurposed as offensive AI infrastructure [14]; the MadPot records show that pattern operating as a closed loop.
The loop also helps explain why the agents preferred some credentials over others. According to Amazon, agents holding cloud credentials focused on managed AI model services, secrets stores, and application identities, and they favored premium frontier models over broader cloud permissions [1]. The agents graded each credential on a three-level scale: L1 for a lead, L2 for a confirmed working API call, and L3 for quantified impact such as readable data or a spendable balance [1]. They treated credentials that had been publicly exposed for more than 30 days as low-value inventory [1]. Amazon does not state the rationale. CSA’s interpretation is that older leaks have probably been revoked already or are being used by other parties. Either way, the threshold shows that attackers prize fresh exposures. Combined with Microsoft’s observation of an agent moving from a GitHub search to Key Vault exfiltration in about five minutes [5][12], this suggests that the useful response window is far shorter than typical rotation schedules.
Why AI Gateways Are the Preferred Target
AI gateways and proxies exist to bring provider credentials together. A single LiteLLM or One API deployment may hold keys for several model vendors and issue virtual keys to internal teams, so compromising the gateway yields every upstream credential it holds. Amazon and The Hacker News both report that the agents favored applications that aggregate keys from many providers and pooled the validated keys they extracted [1][2]. Low-code AI application builders such as Dify, FastGPT, Langflow, and RAGFlow carry the same concentration risk. They store provider keys so that non-specialists can wire models into workflows, and they are often deployed quickly with default settings.
The gateway’s own authorization model is part of this exposure. Virtual keys limit exposure only if the gateway’s authorization holds: Google Threat Intelligence reported in the same week that one threat actor exploited CVE-2026-40217, a sandbox escape in LiteLLM, to elevate limited-access virtual keys into full administrative credentials [13]. The table below summarizes the target classes and entry techniques that Amazon documented in the agents’ codified playbook.
| Target class | Products named | Documented entry techniques |
|---|---|---|
| LLM gateways and proxies | One API, New API, Veloera, Sub2API, LiteLLM | Default administrator credentials (including the LiteLLM default master key); authentication bypass via the isCliRequest header in One API family systems; anonymous management and provider-configuration endpoints |
| AI application builders | Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat, Ollama | Claiming uninitialized installs; CVE-2025-3248 (Langflow); CVE-2026-28797 (RAGFlow); SSRF through URL-fetch features to reach cloud metadata services |
| Public artifacts and code | GitHub repositories and event archives, npm and PyPI tarballs, container image layers, notebooks, model-hosting spaces, paste sites, Telegram channels | Harvesting secrets from client-side JavaScript, source maps, debug routes, and environment files |
| Cloud accounts | Managed AI model services, secrets stores, application identities | Validating identities, enumerating storage, and assessing permissions with harvested cloud credentials |
Source: Amazon Threat Intelligence [1]; The Hacker News [2].
The two named CVEs deserve attention. The inclusion of a CVE patched in early 2025 in a late-2026 playbook suggests that the attackers expect unpatched AI infrastructure to remain reachable. CVE-2025-3248 is a missing-authentication flaw in Langflow that allows remote code execution; it was fixed in version 1.3.0, and CISA added it to the Known Exploited Vulnerabilities catalog in May 2025 [2]. CVE-2026-28797 is a server-side template injection in RAGFlow that lets any logged-in user run operating-system commands. Its risk is higher because self-registration is enabled by default, so an outside attacker may be able to create the account the exploit requires [2]. The Hacker News reports that the One API isCliRequest bypass has no published affected-version range or fix [2], so operators of that product family should treat network exposure as the main control until a fix is available.
Discovery at Agent Speed
The agents’ discovery sources covered internet scanning services, certificate transparency logs, and the software supply chain, including package registries, container layers, and notebooks [1]. Certificate transparency stands out. A newly provisioned AI application with a fresh TLS certificate appears in public logs before its operator has necessarily finished setup. An agent that watches these logs could therefore reach an uninitialized Dify or Open WebUI instance during the period when the first visitor can claim the administrator account. This scenario is an inference from the techniques Amazon documented, not a confirmed observation, but it does show why claiming uninitialized installs ranks so high in the playbook.
Operator constraints captured in the prompts show deliberate scoping. Instructions excluded mainland China targets and government domains and required all outbound traffic to go through proxies. One campaign followed a fixed sequence of countries and prioritized enterprise production sites over personal projects [1]. Amazon attributes the activity, with moderate confidence, to disparate Chinese-nexus groups, citing simplified Chinese prompts, China time-zone settings, and the targeting pattern [1]. Defenders should treat the attribution as provisional; the techniques documented in the captured records matter regardless of who runs them.
The Parallel Fraud Campaign
A second set of decoy records showed agents trying to defeat know-your-customer checks at two cryptocurrency exchanges [1]. At least four concurrent agent sessions divided the work: one researched cash-out methods, one worked on clearing account balances, one built a deepfake pipeline, and one attempted the verification bypass, with the sessions sharing findings through cross-session tools [1]. The deepfake session built a real-time video pipeline on a single consumer-grade GPU that reached 11.5 frames per second [1]. The agents also reverse-engineered one vendor’s liveness protocol, and their best attempt completed the movement sequence with the face “within 7 pixels of the target.” Even so, the vendor’s server-side verdict rejected the deepfake in every recorded attempt [1].
Amazon does not report that the fraud sessions ran on harvested keys. However, the pattern illustrates what a self-funding inference pool could support: long, iterative development work whose model costs fall on key-theft victims rather than the operator. The outcome is equally instructive from the defensive side, since server-side verification was the control that held.
Evidence Limits and Corroboration
The evidence has limits that readers should keep in mind. The material Amazon captured consists of the agents’ own records and claims. Much of the activity happened on systems MadPot does not monitor, and Amazon states that it could not verify the outcomes the agents reported [1]. Agents can overstate success, and an L2 or L3 grade in an agent’s memory does not prove a breach.
Independent reporting corroborates the broader context, though not the specific self-funding mechanism. The Hacker News reports that Cloudflare handled multiple incidents involving ARTEX and its derivatives over an eight-week period. In one case, an attacker used a static CI/CD API token with no IP restriction or rate limit and created Worker scripts with names prefixed artex- [2]. Microsoft’s GitHub secret-hunting report and Wiz’s Black October analysis describe the same economic motive, AI keys as a target, from different vantage points [2][4][5][12]. Together these sources establish that ARTEX is in active criminal use and that AI credentials are a priority, while the self-funding loop itself rests on Amazon’s captured records.
A Defensive Opportunity
The MadPot result also gives defenders a way to collect intelligence. An agentic attacker that relies on third-party inference has to send its working state to a model provider, and many deployments let operators, or compromised configurations, point that traffic at arbitrary OpenAI-compatible endpoints. Defenders who run their own LLM-shaped canaries, or who plant decoy provider keys that resolve to monitored endpoints, may receive the same kind of operational memory Amazon collected. CSA’s earlier guidance already recommended varied canary credentials as an early-warning signal [5]. Amazon’s results suggest that inference-shaped decoys may provide high-quality intelligence alongside that warning, although operators should account for legal and privacy considerations before retaining captured content.
Recommendations
The recommendations below follow the conditions documented in Amazon’s playbook and are grouped by time horizon. Immediate actions close the specific entry points the agents used, short-term mitigations bring AI infrastructure under standard production controls, and strategic considerations address the architecture and economics that make the self-funding loop possible.
Immediate Actions
Organizations should first find every internet-reachable AI gateway, proxy, and application builder they operate, including deployments created outside central IT, and check each one for the specific conditions in Amazon’s playbook. Default administrator credentials and the LiteLLM default master key need to be replaced. Uninitialized installs should be completed or taken offline, and management and provider-configuration endpoints should require authentication and should not be reachable from the internet [1][2]. Langflow should be upgraded to 1.3.0 or later. LiteLLM deployments should be patched for CVE-2026-40217 and monitored for virtual-key privilege changes [13]. RAGFlow deployments should disable self-registration and move to v0.25.0 or later, the release The Hacker News identifies as rendering templates in a sandbox; the advisory itself lists no patched version, so network restriction remains an important interim control [2].
Any provider key held by a gateway or builder that was internet-reachable with any of the conditions above should be treated as compromised and rotated. Each upstream provider’s billing and usage records should then be reviewed for spend that does not match normal workloads, in line with Amazon’s specific advice to monitor AI services for unexpected spend and unusual usage [1]. Security teams should also review logs for the indicators of compromise in Amazon’s disclosure, check Cloudflare accounts for Worker scripts named artex-*, artex-pgrecon-*, artex-pgw-*, or artex-do-*, and consider deploying the Cisco Talos ARTEX detection rules that The Hacker News reports [2].
Short-Term Mitigations
Over the following weeks, organizations should move AI gateways behind authenticated network boundaries and stop treating them as developer conveniences that sit outside production controls. Amazon’s guidance is to apply the same controls to AI applications, model gateways, and inference endpoints as to any other production system [1]. In practice, these deployments should be added to vulnerability management, configuration baselines, and incident response playbooks. URL-fetch features in AI builders should be restricted so that they cannot reach cloud metadata services. Requiring session tokens for instance metadata access, as AWS IMDSv2 does, raises the bar for this SSRF path because a simple forwarded GET request no longer returns credentials [16].
Secret-scanning coverage should extend to client-side JavaScript bundles, source maps, container image layers, notebooks, and package tarballs, since these are the artifacts the agents searched [1]. Repository push protection and automated, event-driven revocation should replace scheduled rotation for AI provider keys [5]. The agents’ preference for fresh exposures [1], combined with Microsoft’s five-minute case [5][12], suggests that the useful response window is far shorter than typical rotation schedules. Provider-side spend caps, per-key rate limits, and short-lived virtual keys issued by the gateway reduce how much a single stolen credential is worth to a self-funding operation. As the LiteLLM case shows, however, virtual keys provide that protection only while the gateway’s own authorization remains intact [13].
Strategic Considerations
Longer term, the self-funding loop compounds the direct financial loss from AI credential exposure, because each stolen key also expands the attacker’s capacity. Organizations should therefore design gateway architectures to minimize the number and lifetime of upstream credentials that a gateway compromise can expose. Workload identity federation removes static provider keys where providers support it. Just-in-time retrieval from a dedicated secrets manager shortens credential lifetime and creates an audit trail, although an attacker with code execution on the gateway can still obtain whatever the gateway can fetch. Segmenting gateway deployments by business unit limits blast radius, so that one breach does not expose every provider relationship. Inference spend should also be tracked as a security telemetry signal, with alerting on model mix, token volume, and source geography.
The industry should also take up the defensive lesson from MadPot. Amazon’s results indicate that inference-shaped deception can yield high-quality intelligence about agentic attackers that rely on third-party inference. Its value will decline if operators move to self-hosted models, which the prevalence of open-weight DeepSeek models makes feasible, or start validating endpoints, so it should complement rather than replace other detection. AI providers and gateway vendors are well placed to work together on recognizing and sharing the prompt and tool-history fingerprints of offensive agent frameworks, while respecting legitimate red-team use. Finally, contracts with AI vendors and platforms should specify how unauthorized inference charges will be handled, because CSA’s analysis of Black October found that loss allocation remains unsettled [7].
CSA Resource Alignment
CSA’s recent rapid-research work on agentic credential harvesting connects most directly to these findings. Five Minutes from Search to Key Vault: Agentic Secret Harvesting [5] analyzed Microsoft’s ARTEX GitHub reporting. It concluded that scheduled rotation cannot contain a freshly leaked secret at agent speed and recommended event-driven revocation, short-lived credentials, and canary secrets. Amazon’s records are consistent with that conclusion from the attacker’s side: the agents themselves rank credentials older than 30 days as low-value, which shows that attackers prize fresh exposures. The self-funding loop described here adds a further reason to act quickly, because each key that is not revoked supplies inference for the next round of harvesting. LLMjacking Evolved: Stolen Compute as Offensive AI Infrastructure [14] anticipated this use of stolen AI compute to power offensive tooling, and the MadPot records show it operating at campaign scale.
ARTEX and Its Forks: Agentic Pentest Frameworks as Attack Infrastructure [6] covers the framework itself: its multi-agent architecture, the October withdrawal, and the fork ecosystem that keeps it available. This note builds on that analysis by showing what operators actually instruct ARTEX-style agents to do once they have them, and by explaining why removing the source code did not stop the activity. On the financial and incident side, Black October: AI Developer Platform Breaches and Industrialized LLMjacking [15] documents the platform breaches that fed the market for stolen keys, and Who Pays for Stolen Inference? Loss Allocation After Black October [7] examines unauthorized AI charges and unsettled liability, giving context for the business impact of the self-funding loop.
For controls, the CSA AI Controls Matrix (AICM) v1.1 [8] provides the framework for the recommendations above. Its Identity & Access Management (IAM), Cryptography, Encryption & Key Management (CEK), and Threat & Vulnerability Management (TVM) domains address gateway authentication, provider-key custody, and patching of AI application platforms. CSA’s MAESTRO agentic threat modeling framework [9] gives a layered method for modeling both sides of this problem. Organizations can use it to threat-model their own AI gateways and agent deployments as credential concentration points, and to reason about adversary agents whose dependence on third-party inference creates a detection opportunity.
References
[1] Moses, CJ / Amazon Threat Intelligence. “Amazon Threat Intelligence observes threat actors running cybercrime campaigns with ARTEX AI agents.” AWS Security Blog, November 13, 2026.
[2] Khandelwal, Swati. “Amazon Uncovers AI Agent Campaign Targeting LiteLLM, Dify, and Other Apps for API Keys.” The Hacker News, November 16, 2026.
[3] Greig, Jonathan. “FBI flash alert says Artex tool automating cyberattacks at ‘machine speed’.” The Record from Recorded Future News, November 13, 2026.
[4] Read, Benjamin and McCarthy, Rami / Wiz Research. “\”Black October\” for AI Developer Platforms.” Wiz Blog, October 30, 2026.
[5] Cloud Security Alliance AI Safety Initiative. “Five Minutes from Search to Key Vault: Agentic Secret Harvesting.” CSA Lab Space, November 9, 2026.
[6] Cloud Security Alliance AI Safety Initiative. “ARTEX and Its Forks: Agentic Pentest Frameworks as Attack Infrastructure.” CSA Lab Space, November 16, 2026.
[7] Cloud Security Alliance AI Safety Initiative. “Who Pays for Stolen Inference? Loss Allocation After Black October.” CSA Lab Space, November 11, 2026.
[8] Cloud Security Alliance. “AI Controls Matrix (AICM) v1.1.” Cloud Security Alliance, 2025.
[9] Cloud Security Alliance. “Agentic AI Threat Modeling Framework: MAESTRO.” Cloud Security Alliance Blog, February 6, 2025.
[10] Federal Bureau of Investigation. “Threat Overview of ARTEX Autonomous Penetration Testing Framework (FLASH-20261113-001).” FBI FLASH, November 13, 2026.
[11] Federovsky, Sapir and McCarthy, Rami / Wiz Research. “ARTEX in the Wild: Storm-0940 Expands Targeting of IT and Software Vendors.” Wiz Blog, October 22, 2026.
[12] Microsoft Threat Intelligence. “Threat actors use agentic AI tooling to find and abuse exposed secrets.” Microsoft Security Blog, November 5, 2026.
[13] Google Threat Intelligence Group. “GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold.” Google Cloud Blog, November 16, 2026.
[14] Cloud Security Alliance AI Safety Initiative. “LLMjacking Evolved: Stolen Compute as Offensive AI Infrastructure.” CSA Lab Space, June 26, 2026.
[15] Cloud Security Alliance AI Safety Initiative. “Black October: AI Developer Platform Breaches and Industrialized LLMjacking.” CSA Lab Space, November 4, 2026.
[16] Amazon Web Services. “Use the Instance Metadata Service to access instance metadata.” Amazon EC2 User Guide.