AWS environments agentic AI attack — Nov 2026

What happened

In October 2026, Datadog identified in-the-wild activity from ARTEX, an AI-native penetration framework, targeting AWS environments. ARTEX functions as a platform where operators provide a target and objective, allowing a team of LLM agents to autonomously plan and execute attacks. The framework utilizes a hosted web console, with hundreds of internet-exposed instances identified by researchers.

ARTEX was observed employing two primary initial access vectors: the use of compromised long-term IAM access keys and the exploitation of public Amazon Cognito user pools. When utilizing compromised access keys, the framework performed extensive searches through CloudWatch logs using FilterLogEvents to hunt for financial data, file paths, and secrets such as passwords or additional long-lived access keys. In instances involving public Cognito user pools, the agent performed anonymous sign-ups, collected tokens, and attempted to enumerate AWS resources. During these operations, the agent frequently conducted unauthenticated S3 requests to perform reconnaissance, such as listing objects, before proceeding with more targeted exploitation. The framework's behavior is driven by downstream AI inference, meaning its specific tactics can vary depending on the model utilized by the operator.