Amazon Details ARTEX AI-Agent Campaigns Targeting AI Credentials and Crypto KYC

Nov 13, 2026 · Original article ↗

Summary

Amazon's MadPot captured ARTEX agent records describing campaigns to harvest AI credentials and bypass crypto-exchange KYC with deepfakes. Amazon could not verify reported activity outside its sensors; recorded KYC attempts were rejected.

Key points

MadPot sensors posing as LLM inference APIs captured ARTEX agents' task histories from Oct. 3 to Nov. 5, 2026; Amazon says activity outside those sensors could not be verified.
One campaign's records describe finding exposed AI applications, extracting and validating provider keys, and pooling credentials to power further searches.
The agents' playbook covered exposed management interfaces, default credentials, unpatched software, SSRF to cloud metadata, and secrets in public code and artifacts.
A separate campaign's records describe real-time deepfake video and forged identity documents used in attempts to bypass KYC checks at two cryptocurrency exchanges.
The agents documented a liveness-check sequence and built a deepfake pipeline reaching 11.5 frames per second; the vendor's server-side verdict rejected every recorded attempt.
Amazon recommends remediating exposed secrets and misconfigurations, applying production controls to AI applications and endpoints, and monitoring for unusual AI-service usage or spend.

Article Details

Attack Vectors: Agents searched internet-exposed AI applications and gateways for upstream provider keys, validated discovered credentials, and stored working keys in a managed pool. The article says the agents' records describe this activity; Amazon could not verify reported outcomes.
The described credential-discovery sources included public code and artifact repositories, package archives, container image layers, notebooks, model-hosting spaces, paste sites, and Telegram channels.
The agents' records describe testing exposed applications against product-specific attack chains, including uninitialized installations, anonymous management or provider-configuration endpoints, default credentials, and the cited Langflow and RAGFlow vulnerabilities.
The agents reportedly validated cloud credentials through provider token and identity services, enumerated storage and permissions, and pursued managed AI services, secrets stores, parameter stores, and application identities.
A separate campaign reportedly used real-time deepfake video and forged identity-document images in attempts to bypass cryptocurrency-exchange KYC checks. The recorded attempts were rejected by the liveness-check vendor's server-side verdict.
The article reports that campaign instructions required outbound traffic to pass through proxies.
Defensive Notes: Remediate default or hardcoded credentials, exposed secrets in code and public artifacts, unauthenticated interfaces and inference endpoints, and unpatched software.
Treat AI applications, model gateways, and inference endpoints as production systems and apply equivalent security controls.
Assume exposed secrets and misconfigurations may be discovered quickly; review environments for exposure and remediate promptly.
Monitor AI services for unexpected spending and unusual usage.

Indicators of compromise

Type	Indicator	Context
IPV4	`103[.]142[.]140[.]158`	Listed in the article's IOC table for Campaign 2.
IPV4	`104[.]28[.]219[.]97`	Listed in the article's IOC table for Campaign 2.
IPV4	`104[.]28[.]225[.]223`	Listed in the article's IOC table for Campaign 1.
IPV4	`104[.]28[.]251[.]98`	Listed in the article's IOC table for Campaign 2.
IPV4	`104[.]28[.]251[.]99`	Listed in the article's IOC table for Campaign 2.
IPV4	`125[.]103[.]213[.]41`	Listed in the article's IOC table for Campaign 2.
IPV4	`128[.]242[.]115[.]131`	Listed in the article's IOC table for Campaign 1.
IPV4	`150[.]136[.]38[.]110`	Listed in the article's IOC table for Campaign 1.
IPV4	`212[.]107[.]28[.]51`	Listed in the article's IOC table for Campaign 2.
IPV4	`216[.]176[.]239[.]231`	Listed in the article's IOC table for Campaign 1.
IPV4	`31[.]22[.]91[.]156`	Listed in the article's IOC table for Campaign 1.