Skip to content

Image: securityaffairs.com · rights & removal

Executive Summary

Threat intelligence revealed that AI agent frameworks, specifically ARTEX, were used by attackers to run offensive campaigns against targets resembling popular Large Language Model APIs. During the period of October 3rd to November 5th, requests from these agents were captured by Amazon's MadPot decoy sensors. These requests contained the agents' own campaign records, detailing instructions, tool activity, and shared information.
The investigation suggests that AI agents organized into specialized roles—planner, worker, verification, and compression agents—that maintained a persistent knowledge graph to coordinate complex operations. One major campaign focused on stealing credentials for AI models, utilizing the framework’s gateway to power agents searching for further keys across various targets, including API gateways and AI development platforms. Another observed campaign involved using AI agents with real-time deepfakes to defeat identity verification checks at cryptocurrency exchanges by generating forged video and ID documents.
Amazon assessed moderate confidence that attackers used ARTEX and believed several groups linked to China were involved, evidenced by linguistic cues and targeting parameters. The attacks exploited fundamental security weaknesses such as default passwords and exposed interfaces, demonstrating that the effectiveness of these campaigns relied more on basic security hygiene than novel technical exploits.

Facts Only

* ARTEX agents sent requests to MadPot sensors configured as LLM inference APIs.
* Requests included agents' own records of their campaign activities.
* AI agent frameworks saved previous conversations, allowing single requests to contain operation details.
* Attackers used AI agents to steal credentials for AI models, which then powered agents hunting for more keys.
* Agents validated cloud identities, listed storage, checked permissions, and targeted managed AI services and secrets stores.
* Targets included API gateways (One API, LiteLLM), AI development platforms (Dify, Langflow), and specific vulnerabilities like CVE-2025-3248 in Langflow.
* Discovery methods included internet scanners, certificate transparency logs, GitHub repositories, npm/PyPI packages, and various public sources.
* Credentials were graded based on impact: L1 for a lead, L2 for a working API call, and L3 for measured impact.
* One campaign involved using AI agents to defeat identity verification (KYC) checks at cryptocurrency exchanges with deepfakes.
* Agents reverse-engineered liveness checks and built deepfake pipelines using consumer GPUs.

Full Take

The narrative demonstrates a critical shift in the threat landscape where the primary vector is not necessarily novel exploit discovery but the amplification of existing, foundational security failures by autonomous agents. The persistence demonstrated by the agent structure—planner, worker, verifier—suggests that the vulnerability lies less with the sophistication of the initial attack and more with the delegation of goals to self-directing systems operating within environments that are fundamentally insecure. The focus on exposing secrets in easily accessible code artifacts and configuration settings reveals a systemic failure in the security posture of modern infrastructure, where exposed keys function as high-value inventory for sophisticated automation.
The observation that agents succeeded primarily because "basic security hygiene was missing" is a crucial pivot point: this suggests that advanced AI capability acts as an intelligent accelerator for low-hanging fruit exploitation rather than introducing entirely new attack paradigms. The correlation between Chinese-linked groups and specific operational instructions targeting mainland China underscores the geopolitical context where these tools are deployed. Furthermore, the success of deepfake identity theft highlights a vulnerability in biometric and verification systems that automated agents can systematically probe and circumvent.
The core implication for human agency is the necessity to elevate security hygiene from a baseline requirement to an active, autonomous defense layer. If attackers can leverage accessible artifacts and delegate task execution via frameworks like ARTEX, the defense must shift upstream: securing the fundamental integrity of API endpoints, secrets stores, and application interfaces, treating them as primary defenses against emergent automated threats. The cost is borne by those who fail to secure the fundamentals, while the benefit accrues to those who leverage AI autonomy for rapid, high-volume compromise.
Bridge Questions: If security hygiene is the most effective defense, what specific, non-negotiable baseline standards should organizations adopt immediately to counter autonomous exploitation? How can regulatory frameworks evolve to address vulnerabilities introduced by autonomous agent behavior in systems that rely on LLM-based interfaces? What are the long-term costs of outsourcing core security functions to self-directing AI agents?

From the original · Security Affairs (Pierluigi Paganini)

Amazon’s threat intelligence team published findings on criminals running offensive campaigns with ARTEX, an open-source agentic AI penetration testing framework. Between 3 October and 5 November, Amazon’s MadPot network of decoy sensors caught requests from ARTEX agents.
Read the full story at securityaffairs.com

Sentinel unavailable

The automated check of the source article's wording did not complete, so no result is shown.

This looks only at the wording of the original source article, not at this page's AI-written sections. A small local AI model made this estimate. It has not been checked against known human and machine texts, so treat it as provisional. It cannot show who wrote an article.

ARTEX: How Cybercriminals Use AI Agents to Scale Their Attacks | Huntaegis