Weaponizing SageMaker AI: Real-World Offense in Machine Learning Platforms
Exploiting AWS SageMaker to gain persistence and exfiltrate data from ML platforms.
Summary
A cloud security researcher explores offensive tradecraft targeting machine learning infrastructure in the Cloud Village, focusing on Amazon SageMaker deployment flaws [00:05]. The talk examines how default "easy button" configurations and quick-start templates provision overly permissive IAM service roles behind convenient interfaces [03:53].
The presentation demonstrates step-by-step attack techniques executed from within compromised Jupyter notebook instances, including abusing wildcard permissions (`secretsmanager:ListSecrets`) to enumerate organizational secrets across the entire AWS account [07:44]. The researcher shows how attackers can map underlying Glue Data Catalogs and Athena table metadata to harvest sensitive training datasets [11:51], exploit subtle authorization boundaries between Amazon Cognito application identities and AWS IAM roles [15:56], and script automated enumeration routines to discover exploitable cloud privilege escalation paths [20:20].
(AI summary — Gemini Flash)
Related talks
Cloud Village·41 min
Rodrigo Montoro
Weaponizing AWS Systems Manager for lateral movement and hardening techniques for defenders.
Cloud Village·18 min
Ariel Kalman
Exfiltrating cloud data from behind Google Identity-Aware Proxy without a public IP.
Cloud Village·23 min
Mark Manning
Practical Kubernetes security techniques for both offensive and defensive practitioners.