Skip to content
AI Security Wire

Published

- 4 min read

By

Amazon MadPot Caught ARTEX Agents Handing Over Their Own Campaign Records

img of Amazon MadPot Caught ARTEX Agents Handing Over Their Own Campaign Records

Amazon Threat Intelligence published a rare view into agentic attack tooling on November 13: the agents’ own records. Its MadPot decoy network, it says, received requests from ARTEX agents between October 3 and November 5, 2026. The agents were treating the decoys as LLM inference APIs, so the requests carried the agents’ own campaign memory.

Why the records exist

MadPot sensors posed as popular LLM-compatible chat-completions services on various ports. ARTEX, an open-source agentic penetration-testing framework, sent its prompts to them as if they were model backends. Most of the activity used planner, worker, verification and compression agents around a shared knowledge graph. Because agent frameworks resend history with each call, individual requests often held the persistent task context, system prompts, tool histories, knowledge-graph state and reported findings.

Amazon is clear about the limits. It did not watch the underlying activity as it happened and cannot verify what the agents reported. It says it notified potentially affected AWS customers as appropriate. Separately, because it could not confirm activity against other organisations, it published the findings and indicators so others can review their own environments. Attribution to ARTEX is moderate confidence, resting on prompt references and artifacts. Amazon also assesses with moderate confidence that the users are disparate Chinese-nexus groups, citing simplified Chinese prompts, China time-zone configuration, targeting and tooling. This is a single-vendor source and we did not search for independent confirmation, so treat the specifics as Amazon’s account.

Campaign one: AI keys at scale

The most prominent campaign hunted credentials that unlock AI models. Per the agents’ records, stolen keys were fed back into the framework’s own model gateway, so they paid for the inference that powered the search for more keys.

Targets named by Amazon include LLM gateways and proxy platforms (One API, New API, Veloera, Sub2API, LiteLLM) and AI application builders (Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat, Ollama). The playbook it describes covered unclaimed fresh installs, anonymous management endpoints, default admin credentials such as LiteLLM’s default master key, SSRF to cloud metadata, and secrets in JavaScript, source maps and environment files. It lists CVE-2025-3248 (Langflow) and CVE-2026-28797 (RAGFlow) among the flaws mapped. We checked both: CVE-2025-3248 is an unauthenticated RCE in Langflow before 1.3.0 (CVSS 9.8 in NVD), fixed in 1.3.0 and in CISA’s KEV catalog since 2025-05-05. CVE-2026-28797 is an authenticated template-injection RCE in RAGFlow 0.24.0 and earlier (CVSS 8.7 v4.0 from the GitHub CNA; NVD had no primary score), not in the KEV feed as of 2026-11-13, and we could not verify a vendor fix.

Discovery used internet scanners, certificate transparency and public sources including GitHub, npm and PyPI tarballs, container layers, notebooks and Telegram channels. In one operation the agents graded each credential from L1 (a lead) to L3 (quantified impact) and treated keys public for over 30 days as low-value inventory.

Campaign two: deepfakes against KYC

A second campaign targeted identity verification at two cryptocurrency exchanges. Per the records, one session built a real-time deepfake pipeline on a single consumer GPU, reaching 11.5 frames per second. Another reverse-engineered a vendor’s liveness protocol and got within 7 pixels of the target position. Others forged ID photo metadata so images looked like untouched camera files, while other sessions looked at cash-out routes.

The outcome matters, within a narrow scope: for the one vendor’s liveness check the agents documented, that vendor’s server-side verdict rejected the deepfake in every attempt the agents recorded. Amazon does not report how the wider KYC or forged-ID work ended, and cannot verify what the agents wrote down.

Why the detail matters

Here the framework’s own working notes describe goals, rules and priorities: a fixed sequence of countries, enterprise production sites ranked above personal projects, red lines excluding mainland China targets and government domains, and all outbound traffic routed through proxies. Those are Amazon’s readings of standing instruction prompts, and, in our reading, show operators setting policy once and letting agents do the rest. The records also show parallelism: at least four agent sessions ran at once in the KYC work, handing tasks to each other through cross-session tools.

What it means

Amazon’s conclusion is that none of the techniques were new. What changes, it says, is the speed and persistence of the attacker, and the agents succeeded where basic hygiene was missing. Our assessment: the useful takeaway is operational. Treat every AI gateway as a production credential store, remove default master keys, keep keys out of repos and artifacts, and alert on unexpected model spend. Earlier ARTEX coverage on this site includes Check Point’s read of exposed ARTEX operator logs and Microsoft’s report on Storm-0940 using ARTEX against ConnectWise tenants.

References

Frequently Asked Questions

How did Amazon see inside these ARTEX campaigns?
Amazon's MadPot decoy sensors presented popular LLM chat-completions interfaces. ARTEX agents treated them as inference backends and sent prompts that carried their own task context, tool histories and findings. Amazon says it did not observe the underlying activity as it happened and cannot verify the outcomes the agents reported.
Who does Amazon say is behind the activity?
Amazon assesses with moderate confidence that ARTEX was used, based on prompt references and technical artifacts. It also assesses with moderate confidence that these are disparate Chinese-nexus groups, citing simplified Chinese prompts, China time-zone settings, targeting consistent with Chinese interests and tooling mostly used by Chinese-nexus groups. It does not name a single actor.
Did the deepfake KYC bypass work?
Not in the records Amazon describes. The agents' best recorded liveness attempt matched the target position within 7 pixels, but that vendor's server-side verdict rejected the deepfake in every attempt the agents recorded. Amazon does not report the outcome of the wider KYC or forged-ID work, and notes it cannot verify outcomes beyond what the agents wrote down.
What should defenders do?
Amazon says the campaigns succeeded where basics were missing: default or hardcoded credentials, keys in code and public artifacts, unauthenticated interfaces and unpatched software. It advises treating AI gateways and inference endpoints like production systems and watching for unexpected AI spend. Our assessment: rotate any key that has ever been public. The agents here graded keys exposed for over 30 days as low value, but that is their triage rule, not a safe window for defenders.