Attack Chain

1. Attacker crafts malicious JSON-Schema: An attacker creates a JSON-Schema file (e.g., probe.json) containing a $ref field with a crafted URI.
2. References sensitive local files: The $ref field uses a file:// scheme to point to an absolute path (e.g., file:///etc/passwd) or ../ path traversal sequences (e.g., ../.ssh/id_rsa) to reference arbitrary local files.
3. Victim executes datamodel-code-generator: A victim's system, such as a CI/CD pipeline, development environment, or multi-tenant service, invokes the datamodel-code-generator utility to process the attacker's untrusted schema (e.g., datamodel-codegen --input probe.json --output o.py).
4. Vulnerable reference resolution: datamodel-code-generator's internal logic, specifically within the _get_ref_body function, attempts to resolve the $ref target.
5. Bypasses security control: The vulnerability allows the file:// scheme to be explicitly exempted from the allow_remote_refs check, or for relative paths, the is_relative_to(base_path) check is not enforced, allowing traversal outside the base directory.
6. Arbitrary file read: The vulnerable code path (_get_ref_body_from_url or _get_ref_body_from_remote) proceeds to load the content of the arbitrary local file referenced in the $ref field.
7. Filesystem oracle and information disclosure: Depending on the target file's content and permissions, the tool either generates an error message (e.g., "TypeError: Expected dict, got str") indicating the file was read and parsed, or directly incorporates schema-shaped secrets (e.g., from const/default/enum/description fields) from the read file into the generated Python code.
8. Exfiltration via generated code or errors: The attacker receives the generated Python code (if it contains secrets) or observes the process's error messages (providing an oracle for file existence and permissions), thereby exfiltrating sensitive information or mapping the victim's filesystem layout.

Impact

This vulnerability leads to arbitrary local file read (CWE-22) and subsequent information disclosure (CWE-200), bypassing a documented security control (--no-allow-remote-refs). Any application, CI pipeline, or multi-tenant service running datamodel-code-generator on untrusted schemas is at risk. Attackers can read any file accessible to the process user, including sensitive system files like /etc/passwd, private keys (e.g., PEM id_rsa), and configuration files (e.g., service-account JSON, Kubernetes secret manifests). This allows for host filesystem mapping by probing arbitrary paths and exfiltrating secret values verbatim if they are embedded within schema-shaped nodes. The issue does not require authentication or special privileges, making it a critical threat for environments processing untrusted input.

Recommendation

• Upgrade pip/datamodel-code-generator to version 0.62.0 or higher to remediate CVE-2026-55389.
• Ensure the --no-allow-remote-refs flag is enabled when invoking datamodel-code-generator with untrusted input, as this flag's bypass for file:// references was fixed in version 0.62.0.
• Restrict execution permissions for datamodel-code-generator processes to the minimum necessary files and directories to limit potential damage from exploitation.
• Block the proof-of-concept URL https://gist.github.com/thegr1ffyn/2a87e81f985883acc30d0118c52da4d3 at network perimeters to prevent access to known exploit code.