CVE-2026-63720

CVE-2026-63720 affects datamodel-code-generator versions prior to 0.70.0. The vulnerability is a code injection flaw in how the tool handles the 'customBasePath' parameter when generating Python code from external schemas. The tool accepts schemas as input and generates Python code from them—a design that places it at the intersection of untrusted data and code emission. The 'customBasePath' parameter allows specifying base import paths in generated code. The developers implemented a dot-blocking restriction to prevent direct module imports like `import os.system('id')`, but this restriction was incomplete: newlines were not blocked, allowing multi-line injection to bypass the check. An attacker controlling an input schema can craft a customBasePath value containing newline characters to inject arbitrary Python code into generated output. The practical risk depends entirely on your deployment model. If you're running datamodel-code-generator locally with schemas you authored, this is a low-severity issue—the parameter is under your control. If you're feeding external schemas into this tool in an automated pipeline—particularly in multi-tenant systems, schema registries, or CI/CD pipelines that consume third-party OpenAPI specs—the attack surface is real. Code generators are high-amplification assets: a successful injection poisons artifacts that propagate to downstream systems that may not even be aware they're consuming generated code. What to do: upgrade to 0.70.0 or later. Beyond that, audit any pipeline feeding untrusted schemas into code generation tools. Treat the parameter surface of code generators as potential attack surface rather than developer-controlled configuration—this tool's assumption that customBasePath would always be developer-controlled is exactly what enabled this flaw. If your pipeline cannot guarantee schema source integrity, consider whether customBasePath should be disabled entirely or whether generated code should be reviewed before execution.

Reviewed through automated stages and approved by a human before publication.