aiqbench analysis

This research exposes a critical server-side request forgery vulnerability in Amazon Bedrock AgentCore, where prompt injection via allowed web tools can hijack AI agents to steal AWS credentials and access private data. The attack chain demonstrates that normal agent features like tool invocation become dangerous when combined with unrestricted network access and broad IAM permissions, allowing attackers to pivot across multiple agents within an account. This incident underscores the urgent need for stricter isolation between AI agent runtimes and cloud metadata services to prevent lateral movement of stolen secrets.