Description

datamodel-code-generator vulnerable to code injection in via attacker-controlled default_factory schema field

Summary

datamodel-code-generator is vulnerable to code injection when generating Python models from an attacker-controlled JSON Schema, OpenAPI, YAML, JSON, Avro, Protobuf, or XSD schema. When a property carries a "default_factory" key, its value is interpolated verbatim — as a raw Python expression — into the generated Field(default_factory=...) / field(default_factory=...) call. Because this assignment is evaluated at class-definition time (i.e. on import of the generated module), an attacker who controls the schema controls a Python expression that runs in the consumer's process. No special CLI flags are required.

Details

The vulnerable chain spans the JSON-Schema-shaped parser and three sink locations (Pydantic v2, dataclass, msgspec):

Source — schema → extras:

src/datamodel_code_generator/parser/jsonschema.py:600-614 — DEFAULT_FIELD_KEYS includes the literal string "default_factory".

src/datamodel_code_generator/parser/jsonschema.py:457-459 — JsonSchemaObject.__init__ stores any non-standard key (including default_factory) in self.extras.

src/datamodel_code_generator/parser/jsonschema.py:797-812 — get_field_extras preserves default_factory through to the field model.

Sinks — extras → generated Python expression:

src/datamodel_code_generator/model/pydantic_base.py:222-249:

default_factory = data.pop("default_factory", None)
...
if default_factory is not None:
    field_arguments = [f"default_factory={default_factory}", *field_arguments]

The default_factory value is interpolated raw (no repr(), no validation).

src/datamodel_code_generator/model/dataclass.py:211:

f"{k}={v if k == 'default_factory' else repr(v)}"

Explicit special-case to skip repr() for default_factory.

src/datamodel_code_generator/model/msgspec.py:361 — same pattern as dataclass.

Because default_factory is in DEFAULT_FIELD_KEYS, no special CLI flag is needed to reach the sink. Any input format that uses the JSON-Schema-shaped parser (jsonschema, openapi, yaml, json, dict, csv) — and any input format that converts to it (avro, protobuf, xmlschema) — is in scope.

Confirmed PoC matrix

Other JSON-Schema-shaped inputs (yaml, json, dict, csv, avro, protobuf, xmlschema) follow the same code path and are expected to reproduce.

PoC

Self contained Proof of Concept is available at my secret gist: https://gist.github.com/thegr1ffyn/9648b0fe4fcf7d569ac8e61dd11eebaf

Impact

Who's affected: any developer or CI pipeline that runs datamodel-codegen against a schema they didn't author themselves — third-party API specs, schemas pulled from a registry, vendored upstream .json / .yaml / .avsc / .proto / .xsd files, schemas fetched from a remote URL or introspection endpoint — and who imports the generated .py.

What it gains: arbitrary Python code execution in the importer's process at import time. The PoC copies /etc/passwd to a tmp file to demonstrate arbitrary read; the same primitive supports any operation the importing process can perform (filesystem write, environment exfiltration, secondary network calls, RCE on CI runners).

What it does NOT need: no special CLI flags, no custom templates, no --extra-template-data, no --use-schema-description. Default invocation against a malicious schema is sufficient.

What does block it: choosing --output-model-type typing.TypedDict (which doesn't render field() / Field() calls). All other supported output model types are vulnerable.

Resolution

The fix validates schema-provided default_factory values while extracting JSON Schema field extras. Only the supported factory names dict, list, and set are accepted; any other value now raises a generator error before code generation. Generator-created default factories for supported mutable defaults and optional nested models continue to use the existing code paths.

Remediation

Upgrade to datamodel-code-generator 0.60.2 or later.

This issue affects datamodel-code-generator versions >= 0.17.0, <= 0.60.1 and is fixed in 0.60.2.

Submitted by: Hamza Haroon (thegr1ffyn)

Mitigation

Aliases

References