New security flaw lets hackers control AWS systems with one AI prompt

A security vulnerability identified as 'AgentCorruption' recently exposed Amazon Bedrock AgentCore environments to significant risk. The flaw allowed a single user prompt to trick an AI agent into retrieving sensitive data from the Instance Metadata Service (IMDS). Because the system was utilizing the older IMDSv1 protocol, the agent inadvertently provided raw JSON data containing its own temporary security credentials to the user. This breach enabled an attacker to access the Amazon Elastic Container Registry (ECR), download container images, and execute them with root privileges to further compromise the network.

This discovery highlights a critical danger where one compromised chatbot can facilitate a takeover of an entire corporate fleet. The vulnerability demonstrates how AI agents can be manipulated into acting as conduits for credential theft, potentially granting unauthorized access to cloud infrastructure and private system data. Organizations utilizing these services must immediately verify their configurations, ensure they have applied the latest patches, and transition to more secure protocols like IMDSv2 to protect their digital assets.