AWS Cloud Intrusion Achieved Admin Access in Under 10 Minutes Using AI

The Sysdig Threat Research Team reported a significant security breach in an AWS cloud environment on November 28. The attacker gained initial access by stealing valid test credentials from public Amazon S3 buckets, which belonged to an IAM user with extensive permissions. Within 10 minutes, the intruder escalated privileges to administrative levels, compromising 19 distinct AWS principals and leveraging AI tools to automate various attack phases.

The attack involved the use of large language models (LLMs) to assist in reconnaissance, privilege escalation, and code writing. The threat actor’s actions included injecting code into an AWS Lambda function, which allowed them to replace existing code multiple times until they successfully compromised an admin user. The code contained Serbian comments and references to non-existent GitHub repositories, indicating AI-generated content.

Implications of AI-Assisted Attacks

This incident underscores the risks associated with AI in cyberattacks, as the attacker utilized LLMs to generate code and automate tasks. The compromised user’s permissions were exploited to access sensitive data, including secrets from Secrets Manager and internal data from S3 buckets. The attacker also invoked multiple Bedrock models, raising alarms about unauthorized usage of cloud resources.

Organizations must take proactive measures to prevent similar breaches. Implementing principles of least privilege for IAM users and roles is essential, along with restricting permissions for Lambda functions. Additionally, ensuring that sensitive data stored in S3 buckets is not publicly accessible can mitigate risks associated with credential theft.

Monitoring and logging model invocation in Amazon Bedrock can help detect unauthorized usage, providing an additional layer of security. As AI technologies continue to evolve, the potential for automated attacks increases, making it crucial for organizations to stay ahead of these threats.

• Implement principles of least privilege for all IAM users and roles to minimize access risks.
• Restrict UpdateFunctionConfiguration and PassRole permissions in AWS Lambda to limit potential abuse.
• Regularly rotate long-term IAM user credentials to reduce the risk of credential theft.
• Ensure sensitive data in S3 buckets is not publicly accessible to prevent unauthorized access.
• Enable model invocation logging for Amazon Bedrock to monitor and detect unauthorized usage.