Fix: Triton inference spec hmac exposure#5654
Open
pravali96 wants to merge 4 commits intoaws:masterfrom
Open
Conversation
added 2 commits
March 19, 2026 20:54
The ONNX export path in _prepare_for_triton() set self.secret_key to a hardcoded value 'dummy secret key for onnx backend'. This key was then passed as SAGEMAKER_SERVE_SECRET_KEY into container environment variables and exposed in plaintext via DescribeModel/DescribeEndpointConfig APIs. The ONNX path does not use pickle serialization — models are exported to .onnx format and loaded natively by Triton's ONNX Runtime backend. There is no serve.pkl, no metadata.json, and no integrity check to perform. The secret key was dead code that also constituted a hardcoded credential (CWE-798). With this change, self.secret_key remains empty string (set by _build_for_triton), and the existing cleanup in _build_for_transformers removes empty SAGEMAKER_SERVE_SECRET_KEY from env_vars before CreateModel. Addresses: P400136088 (Bug 2 - Hardcoded secret key)
Addresses P400136088 Bug 1 and V2146375387 (Triton path).
Three changes:
1. check_integrity.py: Switch from HMAC-SHA256 to plain SHA-256.
- Remove generate_secret_key() — no longer needed
- compute_hash() now uses hashlib.sha256() instead of hmac.new()
- perform_integrity_check() no longer reads SAGEMAKER_SERVE_SECRET_KEY
from environment
2. triton/model.py: Add integrity check in initialize() BEFORE
cloudpickle deserialization. Previously the handler called
cloudpickle.load() with no verification (acknowledged by a TODO
comment). Now reads the file into a buffer, runs
perform_integrity_check(), then deserializes with cloudpickle.loads().
3. triton/server.py: Remove SAGEMAKER_SERVE_SECRET_KEY from container
environment variables in both local and SageMaker deployment modes.
The key is no longer needed since integrity checking uses plain
SHA-256.
4. model_builder_utils.py: Update _hmac_signing() to use plain SHA-256
and stop generating/storing a secret key. Remove generate_secret_key
import.
The integrity check still detects accidental corruption of model
artifacts in S3. The HMAC was providing a false sense of security since
the key was exposed via DescribeModel/DescribeEndpointConfig APIs.
pravali96
force-pushed
the
fix/triton-inference-spec-hmac-exposure
branch
from
3b99611 to
cb365fb
Compare
pravali96
force-pushed
the
fix/triton-inference-spec-hmac-exposure
branch
from
24f9bba to
30d0c21
Compare
pravali96
changed the title
pravali96
force-pushed
the
fix/triton-inference-spec-hmac-exposure
branch
from
9e4d69c to
2db9ae8
Compare
Remove generate_secret_key import and usage from TorchServe, MMS, TF Serving, and SMD prepare functions. Switch compute_hash calls from HMAC-SHA256 to plain SHA-256 (no secret_key parameter). This is required because generate_secret_key was removed from check_integrity.py in the previous commit. Without this change, all model server imports fail with ImportError.
pravali96
force-pushed
the
fix/triton-inference-spec-hmac-exposure
branch
from
2db9ae8 to
8ecfe9d
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Issue #, if available:
Description of changes:
Security fixes for HMAC key exposure and missing integrity check in the Triton inference handler and all model server prepare paths.
Issue
Three security issues were identified in the
sagemaker-servepackage:Missing integrity check in Triton handler (CWE-502):
triton/model.pydeserializedserve.pklviacloudpickle.load()with no integrity verification before execution. A# TODOcomment at line 32 acknowledged this gap. All other handlers (TorchServe, MMS, TF Serving, SMD) already had this check.Hardcoded secret key for ONNX path (CWE-798):
model_builder_utils.pysetself.secret_key = "dummy secret key for onnx backend"in the Triton ONNX export path. This was passed asSAGEMAKER_SERVE_SECRET_KEYinto container environment variables and exposed in plaintext viaDescribeModelAPI. The ONNX path does not use pickle serialization, so no secret key is needed.HMAC secret key exposed via environment variables (CWE-200, CWE-522): All model server implementations injected the HMAC secret key as
SAGEMAKER_SERVE_SECRET_KEYinto container environment variables. These are returned in plaintext byDescribeModel,DescribeEndpointConfig, andDescribeModelPackageAPIs, allowing any principal with read permissions to extract the key and forge valid integrity signatures for malicious pickle payloads.Fix
Switch from HMAC-SHA256 (requires a secret key) to plain SHA-256 (no key needed), matching the approach taken for CVE-2026-1777 (PR #5348/#5379) which made the same change for the remote function path.
Changes
check_integrity.pygenerate_secret_key()— no longer neededcompute_hash()now useshashlib.sha256()instead ofhmac.new()perform_integrity_check()no longer readsSAGEMAKER_SERVE_SECRET_KEYfrom environmentBy submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.