Amazon Detects Campaign Targeting AI Infrastructure and Developer Credentials


Using autonomous artificial intelligence (AI) agents, Amazon Threat Intelligence has identified cybercrime campaigns that identified vulnerable AI applications, stole API keys, and reused the credentials to support further attacks utilizing the same credentials. 

It was likely that credentials from multiple AI providers could be exposed as a result of these operations, which targeted internet-exposed platforms such as LiteLLM, One API, Dify, Langflow, RAGFlow, and Open WebUI.
Using the MadPot network, a globally distributed system of decoy servers designed to monitor suspicious interactions, Amazon's MadPot network provided the basis for these findings.
A request from an AI agent operating through the open-source ARTEX penetration testing framework was sent to the sensors between October 3 and November 5, 2026. These requests contained operational instructions, task histories, and reports of findings, which permitted researchers to examine how the campaigns were conducted.
With moderate confidence, Amazon determined that ARTEX was employed for the operations and that separate China-linked threat groups were involved. 
Based on the evidence, the assessment was made, which included simplified Chinese-language prompts, China-based time zone configurations, targeting patterns, and tooling associated with Chinese-nexus actors. In Amazon's view, the agents operate through a coordinated structure that is comprised of planners, workers, verification agents, and compression agents. 
Each component shared a persistent knowledge graph, which allowed agents to access previous findings, coordinate tasks, and continue operations without repeatedly initiating reconnaissance efforts. Various activities were described in the records, including the discovery of vulnerabilities, the theft of credentials, the expansion of cloud access, and the compromise of cryptocurrency exchange identity verification systems. 
In the primary campaign, API keys were harvested to access large language models (LLMs). Agents searched for internet-based AI applications, extracted provider credentials from those applications, tested whether the credentials remained functional, and added valid credentials to a pool of available credentials. 
It was subsequently determined that the stolen keys could be transferred to the framework's own model gateway, which enabled the agents to use AI services to search for additional credentials.
Due to the fact that compromised applications could expose several sets of credentials, the campaign focused on platforms that manage connections to multiple AI providers. This study identified gateways and proxy services such as LiteLLM, One API, New API, Veloera, and Sub2API as well as artificial intelligence applications such as Dify, FastGPT, Langflow, RAGFlow, LobeChat, Open WebUI, and Ollama as possible targets. 
An attacker prepared a playbook containing identified vulnerabilities associated with these products in order to attack them. Entry points identified included default administrator credentials, incomplete installations, and management interfaces that were not accessible by authentication. There were also references to LiteLLM's default master key, an authentication bypass involving the isCliRequest header in One API family systems, and known Langflow and RAGFlow vulnerabilities. 
To access cloud metadata services, additional techniques included exploiting server-side request forgery (SSRF) through URL-fetching features. Furthermore, the agents looked for exposed credentials in JavaScript, source maps, debugging endpoints, environment files, and publicly accessible code repositories as well. 
Deepfake Campaigns and Cloudflare Incident Reveal Wider ARTEX Threat
The company has also identified an attempt by ARTEX agents to bypass Know Your Customer (KYC) verification systems at two cryptocurrency exchanges in a separate campaign. To avoid detection, the operation involved generating real-time deepfake videos to pass live selfie checks, and modifying identity documents to avoid detection. 
According to reports, agents examined one verification provider's protocol for checking liveness, including instructions for tracking facial features and moving.
However, the provider's server-side verification system rejected every deepfake attempt recorded by Amazon, even though the face was positioned within seven pixels of the target.  Meanwhile, the agents developed a video-generation pipeline capable of producing 11.5 frames per second through the use of a consumer grade graphics processor.
Many agent sessions were involved in the operation, which included bypassing identity checks, preparing deepfake videos, clearing account balances, and identifying methods of cashing out. However, the recorded attempts did not defeat the provider's liveness tests, but they demonstrated that artificial intelligence agents may be able to coordinate multiple stages of a fraud operation with minimal human involvement. 
Cloud environments are also

[…]
Content was trimmed to protect the source. Please visit the original article for the full text.

This article has been indexed from CySecurity News – Latest Information Security and Hacking Incidents