New AI-Powered Threat Allows Hackers to Gain AWS Admin Access in Minutes

Cybersecurity: AI-Driven Cloud Intrusion

A recent sophisticated cyberattack targeted an AWS environment, utilizing AI automation to achieve administrative control in less than 10 minutes. The attack commenced with the discovery of exposed credentials in public Simple Storage Service (S3) buckets, which were named using AI-related terminology.

The attackers employed a method known as LLMjacking, leveraging Large Language Models (LLMs) for automated reconnaissance and malicious code generation. The Sysdig Threat Research Team (TRT) documented this incident, noting the rapid escalation from initial access to administrative privileges.

During the attack, Python code was injected into an existing AWS Lambda function. This script, featuring Serbian comments, enabled the attackers to list all Identity and Access Management (IAM) users and create new administrative access keys.

Lateral Movement and Resource Exploitation

After securing administrative access, the attackers conducted lateral movements across 19 AWS principals, using a combination of compromised accounts and assumed roles. A Terraform module named terraform-bedrock-deploy.tf was identified, designed to deploy a backdoor Lambda function for generating Bedrock credentials.

The primary objective was resource theft, with attackers exploiting Amazon Bedrock to execute multiple high-end AI models, effectively stealing compute time. They also attempted to launch expensive GPU instances, successfully deploying a p4d.24xlarge instance, potentially costing the victim approximately $23,600 per month.

A JupyterLab server was installed on this instance, providing a backdoor for continuous access to computing resources even if AWS credentials were revoked.

This incident underscores the threat posed by AI-driven attacks capable of executing complex operations swiftly. The following mitigation strategies are recommended:

Eliminate Long-Term Credentials: Use IAM roles with temporary credentials instead of static keys. Secure S3 Buckets: Ensure that S3 buckets containing sensitive data are not publicly accessible. Restrict Lambda Permissions: Apply the principle of least privilege, avoiding broad permissions such as UpdateFunctionCode unless necessary. Monitor Bedrock Usage: Set up alerts for unexpected AI model invocations.

Automating detection processes is essential to counter the reduced response time due to AI-driven attack chains.

Based on reporting by GBHackers.