Decoy Model Endpoints Caught ARTEX Agents Narrating Their Own Key-Theft and Deepfake KYC Campaigns

Amazon Threat Intelligence says its MadPot sensors, posing as LLM inference APIs, captured the working memory of AI agents built on the open-source ARTEX pentest framework between October 3 and November 5, 2026, exposing an AI-credential theft operation and deepfake attempts against two crypto exchanges' identity checks.

In a dark warehouse, mechanical moths swarm a decoy amber lantern inside a glass enclosure while a lone observer takes notes, in midnight navy and silver.
In a dark warehouse, mechanical moths swarm a decoy amber lantern inside a glass enclosure while a lone observer takes notes, in midnight navy and silver.

Amazon Threat Intelligence has published a detailed look inside criminal operations run by AI agents, and the evidence came from the agents themselves. In a post dated November 13, 2026, Amazon CISO CJ Moses said that between October 3 and November 5 the company’s MadPot decoy network captured requests from agents built on ARTEX, which Amazon describes as an open-source agentic AI penetration testing framework. The agents mistook MadPot sensors for real LLM inference APIs and sent them their prompts, carrying their own running records of the campaigns they were working on.

How a honeypot ended up reading the playbook

MadPot is Amazon’s globally distributed network of decoy threat sensors. For this work, some sensors were set up to look like popular LLM-compatible chat-completions interfaces on various ports. Agent frameworks that found them treated them as model backends and sent the context they normally pass to a model before making a decision.

That context turned out to be rich. Amazon says the captured requests included persistent task context, system prompts, tool histories, knowledge-graph state and reported findings. Because agent frameworks keep the history of past conversations, a single request often held the full operational history of a campaign, including activity on systems MadPot does not monitor. Amazon notes it did not watch that activity happen and cannot verify the outcomes the agents reported, but the records show what the operators were after.

Most of the activity split work across planner, worker, verification and compression agents that read from and wrote to a shared knowledge graph, so each agent could pick up where another left off.

The largest campaign targeted credentials for AI models. According to Amazon, the agents’ records show stolen keys being fed into the framework’s own model gateway, where they paid for the inference that powered the hunt for more keys. Cloud credentials were a route to more access, though premium frontier model access ranked above general cloud permissions.

The agents favoured software that aggregates provider keys, because one weak admin panel can expose many downstream accounts. Named targets included LLM gateways and proxies such as One API, New API, Veloera, Sub2API and LiteLLM, and app builders such as Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat and Ollama. A playbook mapped each product to known attack paths, including unclaimed fresh installs, anonymous management endpoints, default admin credentials such as LiteLLM’s default master key, CVE-2025-3248 in Langflow, CVE-2026-28797 in RAGFlow, an isCliRequest header bypass in One API-family systems, SSRF to cloud metadata services, and secrets left in JavaScript bundles, source maps and environment files.

Discovery drew on internet scanners, certificate transparency logs, public code and package registries, container images, paste sites and Telegram channels. One operation worked through a fixed list of countries, preferred enterprise production sites, and graded every credential: L1 for a lead, L2 for a confirmed working call and L3 for quantified impact such as readable data or spendable balance. Keys exposed for more than 30 days were filed as low-value inventory.

Deepfakes against liveness checks

A second campaign tried to beat know-your-customer checks at two cryptocurrency exchanges. The records describe agents reverse-engineering one vendor’s liveness protocol, including its face-detection event stream and on-screen coordinate system. The best recorded attempt completed the full head-movement sequence with the face within 7 pixels of the target, yet the vendor’s server-side verdict rejected the deepfake in every attempt the agents logged.

In parallel, a session built a real-time deepfake video pipeline on a single consumer GPU using an open-source face animation model, reaching 11.5 frames per second. Other work forged ID photos with synthetic camera metadata and preserved the original JPEG compression tables so edits would look like untouched camera captures. At least four agent sessions ran at once on cash-out routes, balance clearing, the video pipeline and the verification bypass, handing tasks between them.

Attribution and what defenders should take from it

Amazon assesses with moderate confidence that ARTEX was used, based on prompt references to its tooling, and with moderate confidence that the operators were disparate Chinese-nexus groups, citing simplified Chinese prompts, China time-zone settings, targeting patterns and tooling. Operator instructions excluded mainland China and government domains and required all traffic to pass through proxies. Amazon says it notified potentially affected AWS customers and published 11 IPv4 indicators seen between October 14 and November 2.

The company’s conclusion is that none of the techniques were new. What changed was the speed and persistence of the attacker, and the agents succeeded where basics were missing: default credentials, keys in public code, unauthenticated inference endpoints and unpatched software. Amazon urges organisations to bring AI gateways and model endpoints under the same controls as other production systems and to watch for unexpected AI spend.

Topics
  • #Amazon
  • #Threat Intelligence
  • #AI Agents
  • #Credential Theft
  • #Deepfakes
Raj M

Author

Raj M

Contributor

AI Systems Architect is a seasoned technology leader with over 15 years of experience in the IT industry working with Fortune 500 companies. With a solid foundation in multi-agent systems, open-source LLM infrastructure, and enterprise deployment, he excels at building scalable production-grade AI platforms.