Summary
datamodel-code-generator honours a custom x-python-type JSON-Schema extension that lets a schema author override the generated Python type for a field. The value is forwarded verbatim into the generated Python source as the field annotation, with a single sanitisation pass that is trivial to bypass. An attacker who controls a JSON Schema fed to datamodel-codegen can therefore embed an arbitrary Python statement in the generated module, which executes at class-definition time the moment the developer imports the file. No --extra-template-data and no special flags are required; the vulnerable code is reachable with default settings.
Details
Sink: src/datamodel_code_generator/parser/jsonschema.py, _get_python_type_override (lines 2055–2096, at tag 0.60.1 / commit a321547e):
def _get_python_type_override(self, obj: JsonSchemaObject) -> DataType | None:
x_python_type = obj.extras.get("x-python-type")
if not x_python_type or not isinstance(x_python_type, str):
return None
schema_type = obj.type if isinstance(obj.type, str) else None
if self._is_compatible_python_type(schema_type, x_python_type):
return None
base_type = self._get_python_type_base(x_python_type)
import_ = self._resolve_type_import(base_type)
type_str = x_python_type
prefix = x_python_type.split("[", maxsplit=1)[0]
if "." in prefix: # only sanitiser
type_str = base_type + x_python_type[len(prefix):]
...
...
result = self.data_type(type=type_str, import_=import_)
...
return resultDataType.type flows unescaped into {{ field.type_hint }} in every model template
(model/template/pydantic_v2/BaseModel.jinja2,
model/template/dataclass.jinja2,
model/template/TypedDictClass.jinja2,
model/template/msgspec.jinja2, …).
The only sanitiser — the dot-rewrite at the marked line — fires only when . is in the substring before the first [ in the value. Placing [ early (e.g. X[1]; <payload>) keeps prefix == "X" so the rewrite is skipped and the whole value lands in the generated annotation.
from __future__ import annotations (emitted by default) makes the X[1] portion a lazy string, so X does not need to resolve at runtime. Everything after ; is parsed as a real statement in the class body and is executed when the class is constructed during import.
Output-model types confirmed vulnerable in testing: pydantic_v2.BaseModel, dataclasses.dataclass, typing.TypedDict. msgspec.Struct emits structurally identical code.
PoC
A self-contained PoC is available at: https://gist.github.com/thegr1ffyn/1a7ff2561a581074c49785230b2c5700
Impact
Arbitrary code execution in the developer's interpreter / CI runner as soon as the generated module is imported. Reachable from any workflow that ingests an untrusted JSON Schema:
- OpenAPI / JSON-Schema documents fetched from third-party services or public registries.
- Customer-supplied schemas in B2B platforms that auto-generate client SDKs from user input.
- Schema files added by a malicious commit in a polyglot repository that triggers CI code generation.
The compromise is silent: the schema is valid JSON, the generator emits syntactically clean Python (the trojan statement is a single indented line in the class body), and only the use of the generated file triggers the payload.
Anyone running datamodel-codegen against an attacker-supplied schema is impacted. CI runners and developer workstations are the primary blast radius.
Resolution
The fix validates x-python-type before constructing the generated type annotation. The value is parsed with ast.parse(..., mode="eval") and accepted only when the AST is shaped like a Python type annotation, including names, attributes, subscripts, tuple/list annotation arguments, | unions, and safe literal values where annotation syntax allows them. Statements, calls, and other executable expressions are rejected before code generation. The validator is cached to avoid repeated AST parsing for repeated values.
Remediation
Upgrade to datamodel-code-generator 0.60.2 or later.
This issue affects datamodel-code-generator versions >= 0.51.0, <= 0.60.1 and is fixed in 0.60.2.
Submitted by: Hamza Haroon (thegr1ffyn)
References
thegr1ffyn Reporter
Summary
datamodel-code-generatorhonours a customx-python-typeJSON-Schema extension that lets a schema author override the generated Python type for a field. The value is forwarded verbatim into the generated Python source as the field annotation, with a single sanitisation pass that is trivial to bypass. An attacker who controls a JSON Schema fed todatamodel-codegencan therefore embed an arbitrary Python statement in the generated module, which executes at class-definition time the moment the developer imports the file. No--extra-template-dataand no special flags are required; the vulnerable code is reachable with default settings.Details
Sink:
src/datamodel_code_generator/parser/jsonschema.py,_get_python_type_override(lines 2055–2096, at tag0.60.1/ commita321547e):DataType.typeflows unescaped into{{ field.type_hint }}in every model template(
model/template/pydantic_v2/BaseModel.jinja2,model/template/dataclass.jinja2,model/template/TypedDictClass.jinja2,model/template/msgspec.jinja2, …).The only sanitiser — the dot-rewrite at the marked line — fires only when
.is in the substring before the first[in the value. Placing[early (e.g.X[1]; <payload>) keepsprefix == "X"so the rewrite is skipped and the whole value lands in the generated annotation.from __future__ import annotations(emitted by default) makes theX[1]portion a lazy string, soXdoes not need to resolve at runtime. Everything after;is parsed as a real statement in the class body and is executed when the class is constructed duringimport.Output-model types confirmed vulnerable in testing:
pydantic_v2.BaseModel,dataclasses.dataclass,typing.TypedDict.msgspec.Structemits structurally identical code.PoC
A self-contained PoC is available at: https://gist.github.com/thegr1ffyn/1a7ff2561a581074c49785230b2c5700
Impact
Arbitrary code execution in the developer's interpreter / CI runner as soon as the generated module is imported. Reachable from any workflow that ingests an untrusted JSON Schema:
The compromise is silent: the schema is valid JSON, the generator emits syntactically clean Python (the trojan statement is a single indented line in the class body), and only the use of the generated file triggers the payload.
Anyone running
datamodel-codegenagainst an attacker-supplied schema is impacted. CI runners and developer workstations are the primary blast radius.Resolution
The fix validates
x-python-typebefore constructing the generated type annotation. The value is parsed withast.parse(..., mode="eval")and accepted only when the AST is shaped like a Python type annotation, including names, attributes, subscripts, tuple/list annotation arguments,|unions, and safe literal values where annotation syntax allows them. Statements, calls, and other executable expressions are rejected before code generation. The validator is cached to avoid repeated AST parsing for repeated values.Remediation
Upgrade to
datamodel-code-generator0.60.2or later.This issue affects
datamodel-code-generatorversions>= 0.51.0, <= 0.60.1and is fixed in0.60.2.Submitted by: Hamza Haroon (thegr1ffyn)
References