Add validation for negative position_ids in EmbedLayerNorm

[vraspar]

Description

Adds a missing lower-bound validation check for position_col_index (derived from position_ids input) in the EmbedLayerNormalization contrib operator. A negative value in position_ids could bypass the existing upper-bound-only check and cause an out-of-bounds heap read before the position_embedding buffer.

Motivation and Context

[Copilot]

Pull request overview

This PR adds a missing lower-bound validation check for position_col_index (derived from position_ids input) in the EmbedLayerNormalization CPU contrib op, preventing an out-of-bounds heap read when a negative value is passed as a position ID. A regression test is also added to verify the fix.

Changes:

• Adds position_col_index < 0 guard in embed_layer_norm.cc (the non-quantized CPU path where position_ids is user-supplied data, fixing a real OOB read vulnerability)
• Adds the same position_col_index < 0 guard in qembed_layer_norm.cc (the quantized CPU path, where this check is dead code since position_col_index = index % sequence_length and index is always non-negative)
• Adds a regression test EmbedLayerNormNegativePositionIds in the test file

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 1 comment.

File	Description
onnxruntime/contrib_ops/cpu/bert/embed_layer_norm.cc	Adds < 0 lower-bound check to prevent OOB heap read on negative position_ids — this is the real fix
onnxruntime/contrib_ops/cpu/quantization/qembed_layer_norm.cc	Adds < 0 check to position_col_index, which is dead code since the value is always index % sequence_length (non-negative)
onnxruntime/test/contrib_ops/embed_layer_norm_op_test.cc	Regression test for the negative position_ids validation fix

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[github-actions]

You can commit the suggested changes from lintrunner.

[yuslepukhin]

Make sure to comment on the Copilot comments and resolve them

[yuslepukhin]

LGTM