fix: Security fixes for Triton HMAC key exposure and missing integrity check (v2)

Backport of v3 security fixes for P400136088 and V2146375387.

1. check_integrity.py: Switch from HMAC-SHA256 to plain SHA-256.
   Remove generate_secret_key, remove env var dependency.

2. triton/model.py: Add integrity check in initialize() BEFORE
   cloudpickle deserialization.

3. triton/server.py: Remove SAGEMAKER_SERVE_SECRET_KEY from
   container environment variables.

4. triton/triton_builder.py: Remove hardcoded dummy secret key
   for ONNX path. Rename _hmac_signing to _compute_integrity_hash.
   Use plain SHA-256.

5. All prepare.py files (torchserve, mms, tf_serving, smd):
   Remove generate_secret_key usage, switch to plain SHA-256.

Author: Pravali Uppugunduri

src/sagemaker/serve/model_server/multi_model_server/prepare.py

@@ -26,7 +26,6 @@
 from sagemaker.serve.spec.inference_spec import InferenceSpec
 from sagemaker.serve.detector.dependency_manager import capture_dependencies
 from sagemaker.serve.validations.check_integrity import (
-    generate_secret_key,
     compute_hash,
 )
 from sagemaker.remote_function.core.serialization import _MetaData
@@ -120,11 +119,10 @@ def prepare_for_mms(
 
     capture_dependencies(dependencies=dependencies, work_dir=code_dir)
 
-    secret_key = generate_secret_key()
     with open(str(code_dir.joinpath("serve.pkl")), "rb") as f:
         buffer = f.read()
-    hash_value = compute_hash(buffer=buffer, secret_key=secret_key)
+    hash_value = compute_hash(buffer=buffer)
     with open(str(code_dir.joinpath("metadata.json")), "wb") as metadata:
         metadata.write(_MetaData(hash_value).to_json())
 
-    return secret_key
+    return ""

src/sagemaker/serve/model_server/smd/prepare.py

@@ -12,7 +12,6 @@
 from sagemaker.serve.spec.inference_spec import InferenceSpec
 from sagemaker.serve.detector.dependency_manager import capture_dependencies
 from sagemaker.serve.validations.check_integrity import (
-    generate_secret_key,
     compute_hash,
 )
 from sagemaker.remote_function.core.serialization import _MetaData
@@ -64,11 +63,10 @@ def prepare_for_smd(
 
     capture_dependencies(dependencies=dependencies, work_dir=code_dir)
 
-    secret_key = generate_secret_key()
     with open(str(code_dir.joinpath("serve.pkl")), "rb") as f:
         buffer = f.read()
-    hash_value = compute_hash(buffer=buffer, secret_key=secret_key)
+    hash_value = compute_hash(buffer=buffer)
     with open(str(code_dir.joinpath("metadata.json")), "wb") as metadata:
         metadata.write(_MetaData(hash_value).to_json())
 
-    return secret_key
+    return ""

src/sagemaker/serve/model_server/tensorflow_serving/prepare.py

@@ -11,7 +11,6 @@
 )
 from sagemaker.serve.detector.dependency_manager import capture_dependencies
 from sagemaker.serve.validations.check_integrity import (
-    generate_secret_key,
     compute_hash,
 )
 from sagemaker.remote_function.core.serialization import _MetaData
@@ -57,11 +56,10 @@ def prepare_for_tf_serving(
         raise ValueError("SavedModel is not found for Tensorflow or Keras flavor.")
     _move_contents(src_dir=mlflow_saved_model_dir, dest_dir=saved_model_bundle_dir)
 
-    secret_key = generate_secret_key()
     with open(str(code_dir.joinpath("serve.pkl")), "rb") as f:
         buffer = f.read()
-    hash_value = compute_hash(buffer=buffer, secret_key=secret_key)
+    hash_value = compute_hash(buffer=buffer)
     with open(str(code_dir.joinpath("metadata.json")), "wb") as metadata:
         metadata.write(_MetaData(hash_value).to_json())
 
-    return secret_key
+    return ""

src/sagemaker/serve/model_server/torchserve/prepare.py

@@ -13,7 +13,6 @@
 from sagemaker.serve.spec.inference_spec import InferenceSpec
 from sagemaker.serve.detector.dependency_manager import capture_dependencies
 from sagemaker.serve.validations.check_integrity import (
-    generate_secret_key,
     compute_hash,
 )
 from sagemaker.serve.validations.check_image_uri import is_1p_image_uri
@@ -69,11 +68,10 @@ def prepare_for_torchserve(
 
     capture_dependencies(dependencies=dependencies, work_dir=code_dir)
 
-    secret_key = generate_secret_key()
     with open(str(code_dir.joinpath("serve.pkl")), "rb") as f:
         buffer = f.read()
-    hash_value = compute_hash(buffer=buffer, secret_key=secret_key)
+    hash_value = compute_hash(buffer=buffer)
     with open(str(code_dir.joinpath("metadata.json")), "wb") as metadata:
         metadata.write(_MetaData(hash_value).to_json())
 
-    return secret_key
+    return ""

src/sagemaker/serve/model_server/triton/model.py

@@ -26,10 +26,12 @@ def auto_complete_config(auto_complete_model_config):
     def initialize(self, args: dict) -> None:
         """Placeholder docstring"""
         serve_path = Path(TRITON_MODEL_DIR).joinpath("serve.pkl")
+        metadata_path = Path(TRITON_MODEL_DIR).joinpath("metadata.json")
         with open(str(serve_path), mode="rb") as f:
-            inference_spec, schema_builder = cloudpickle.load(f)
+            buffer = f.read()
 
-        # TODO: HMAC signing for integrity check
+        perform_integrity_check(buffer=buffer, metadata_path=metadata_path)
+        inference_spec, schema_builder = cloudpickle.loads(buffer)
 
         self.inference_spec = inference_spec
         self.schema_builder = schema_builder

src/sagemaker/serve/model_server/triton/server.py

@@ -43,7 +43,6 @@ def _start_triton_server(
         env_vars.update(
             {
                 "TRITON_MODEL_DIR": "/models/model",
-                "SAGEMAKER_SERVE_SECRET_KEY": secret_key,
                 "LOCAL_PYTHON": platform.python_version(),
             }
         )
@@ -146,7 +145,6 @@ def _upload_triton_artifacts(
         env_vars = {
             "SAGEMAKER_TRITON_DEFAULT_MODEL_NAME": "model",
             "TRITON_MODEL_DIR": "/opt/ml/model/model",
-            "SAGEMAKER_SERVE_SECRET_KEY": secret_key,
             "LOCAL_PYTHON": platform.python_version(),
         }
         return s3_upload_path, env_vars

src/sagemaker/serve/model_server/triton/triton_builder.py

@@ -23,7 +23,6 @@
 from sagemaker.serve.detector.pickler import save_pkl
 from sagemaker.serve.model_server.triton.config_template import CONFIG_TEMPLATE
 from sagemaker.serve.validations.check_integrity import (
-    generate_secret_key,
     compute_hash,
 )
 
@@ -213,7 +212,7 @@ def _prepare_for_triton(self):
             export_path.mkdir(parents=True)
 
         if self.model:
-            self.secret_key = "dummy secret key for onnx backend"
+            # ONNX path: no pickle serialization, no serve.pkl, no integrity check needed.
 
             if self._framework == "pytorch":
                 self._export_pytorch_to_onnx(
@@ -237,26 +236,23 @@ def _prepare_for_triton(self):
 
             self._pack_conda_env(pkl_path=pkl_path)
 
-            self._hmac_signing()
+            self._compute_integrity_hash()
 
             return
 
         raise ValueError("Either model or inference_spec should be provided to ModelBuilder.")
 
-    def _hmac_signing(self):
-        """Perform HMAC signing on picke file for integrity check"""
-        secret_key = generate_secret_key()
+    def _compute_integrity_hash(self):
+        """Compute SHA-256 integrity hash on pickle file for integrity check"""
         pkl_path = Path(self.model_path).joinpath("model_repository").joinpath("model")
 
         with open(str(pkl_path.joinpath("serve.pkl")), "rb") as f:
             buffer = f.read()
-        hash_value = compute_hash(buffer=buffer, secret_key=secret_key)
+        hash_value = compute_hash(buffer=buffer)
 
         with open(str(pkl_path.joinpath("metadata.json")), "wb") as metadata:
             metadata.write(_MetaData(hash_value).to_json())
 
-        self.secret_key = secret_key
-
     def _generate_config_pbtxt(self, pkl_path: Path):
         config_path = pkl_path.joinpath("config.pbtxt")
 

src/sagemaker/serve/validations/check_integrity.py

@@ -1,29 +1,21 @@
-"""Validates the integrity of pickled file with HMAC signing."""
+"""Validates the integrity of pickled file with SHA-256 hash."""
 
 from __future__ import absolute_import
-import secrets
 import hmac
 import hashlib
-import os
 from pathlib import Path
 
 from sagemaker.remote_function.core.serialization import _MetaData
 
 
-def generate_secret_key(nbytes: int = 32) -> str:
-    """Generates secret key"""
-    return secrets.token_hex(nbytes)
-
-
-def compute_hash(buffer: bytes, secret_key: str) -> str:
-    """Compute hash value using HMAC"""
-    return hmac.new(secret_key.encode(), msg=buffer, digestmod=hashlib.sha256).hexdigest()
+def compute_hash(buffer: bytes) -> str:
+    """Compute SHA-256 hash of the given buffer."""
+    return hashlib.sha256(buffer).hexdigest()
 
 
 def perform_integrity_check(buffer: bytes, metadata_path: Path):
-    """Validates the integrity of bytes by comparing the hash value"""
-    secret_key = os.environ.get("SAGEMAKER_SERVE_SECRET_KEY")
-    actual_hash_value = compute_hash(buffer=buffer, secret_key=secret_key)
+    """Validates the integrity of bytes by comparing the hash value."""
+    actual_hash_value = compute_hash(buffer=buffer)
 
     if not Path.exists(metadata_path):
         raise ValueError("Path to metadata.json does not exist")

tests/unit/sagemaker/serve/model_server/multi_model_server/test_multi_model_server_prepare.py

@@ -68,7 +68,6 @@ def test_start_invoke_destroy_local_multi_model_server(self):
                 "KEY": "VALUE",
                 "SAGEMAKER_SUBMIT_DIRECTORY": "/opt/ml/model/code",
                 "SAGEMAKER_PROGRAM": "inference.py",
-                "SAGEMAKER_SERVE_SECRET_KEY": "secret_key",
                 "LOCAL_PYTHON": platform.python_version(),
             },
         )

tests/unit/sagemaker/serve/model_server/tei/test_server.py

@@ -68,7 +68,6 @@ def test_start_invoke_destroy_local_tei_server(self, mock_requests):
                 "HF_HOME": "/opt/ml/model/",
                 "HUGGINGFACE_HUB_CACHE": "/opt/ml/model/",
                 "KEY": "VALUE",
-                "SAGEMAKER_SERVE_SECRET_KEY": "secret_key",
             },
         )