Black Hat 2024: Intent Guard Tool Secures LLM-Generated IaC Configurations Against Semantic Attacks

At Black Hat 2024, Chris Wysopal (co-founder of Veracode) and Anna Bakha (CTO of Jerona Gamba) presented Intent Guard, an open-source tool designed to secure LLM-generated infrastructure-as-code (IaC) configurations by detecting semantic misalignments and intent violations. The talk focused on risks posed by untrusted IaC templates—such as CloudFormation, Terraform, and Kubernetes manifests—where attackers embed malicious intent via comments, variable names, or defaults, evading traditional IaC scanners like Checkov. Intent Guard addresses this by modeling infrastructure semantics into formal mathematical representations, enabling intent validation across four stages: context retrieval, infrastructure encoding, intent inference, and semantic validation. The tool identified 172 misconfigurations in 45 real-world projects, including privilege escalations and unintended network exposures, which static scanners missed due to their inability to analyze execution paths or multi-hop privilege chains. Key takeaways emphasized defining explicit security intent, validating behavior (not just configuration), treating all internet-sourced templates as untrusted, and integrating intent validation into CI/CD pipelines. The open-source demo version, released at the presentation, currently supports AWS but is designed to extend to other platforms.