How LLM-Generated Access Policies Can Quietly Expand Permissions

Summary

Researcher Vatsal Gupta says LLM-written access policies can compile yet omit constraints, deny rules, or context, expanding permissions. He recommends validation, behavioral testing, and explicit deny-by-default controls.

Key points

• Gupta says LLM-generated policies can be syntactically valid but semantically wrong, potentially granting access beyond what was intended.
• Examples include omitted region, department, or ownership limits and missing deny rules.
• Policies may rely on nonexistent or incorrectly mapped attributes, or drop time, approval, and session conditions.
• Misclassifying an action can broaden permissions, such as allowing a different operation than the one intended.
• Gupta recommends validating generated policies against expected behavior, testing them rather than only compiling them, and explicitly enforcing deny-by-default.

People