CVE-2026-10800: PaddlePaddle FastDeploy Weak Hash Flaw

CVE-2026-10800 Overview

CVE-2026-10800 affects PaddlePaddle FastDeploy versions up to 2.4.1. The weakness resides in the hash_features function in fastdeploy/multimodal/hasher.py, part of the MultimodalHasher component. The flaw is classified as use of a weak hash algorithm [CWE-327]. An attacker with local access can manipulate inputs handled by the hashing routine to undermine its integrity guarantees. Exploitation requires high attack complexity, and the issue is documented as difficult to exploit in practice. The maintainers have published commit 374945747652a8d32965591c0c01a00c88b7067f to address the issue.

Critical Impact
Use of a weak hash function in the multimodal feature hasher can lead to collisions or predictable digests, undermining integrity assumptions in downstream model serving workflows.

Affected Products

PaddlePaddle FastDeploy up to and including version 2.4.1
fastdeploy/multimodal/hasher.py module
MultimodalHasher component (hash_features function)

Discovery Timeline

2026-06-04 - CVE-2026-10800 published to NVD
2026-06-04 - Last updated in NVD database

Technical Details for CVE-2026-10800

Vulnerability Analysis

The vulnerability is a weak hash algorithm issue [CWE-327] in PaddlePaddle FastDeploy, an inference and deployment toolkit used with multimodal models. The hash_features function in fastdeploy/multimodal/hasher.py uses a cryptographically weak hash to derive identifiers or integrity values for multimodal feature inputs. Weak hashes can produce predictable outputs and collisions, allowing an actor to craft inputs that hash to the same digest as legitimate data.

Root Cause

The root cause is the selection of an unsuitable hash primitive within the MultimodalHasher component. The function relies on a non-cryptographic or deprecated hash for purposes that require collision resistance. Applications that trust the resulting digest for caching, deduplication, or integrity checks inherit the weakness from FastDeploy.

Attack Vector

Exploitation requires local access to the system running FastDeploy and the ability to influence inputs processed by hash_features. The attack complexity is high because the attacker must engineer inputs whose hashes collide with target values and operate within the FastDeploy execution context. The CVE record states that exploitation is known to be difficult.

No public proof-of-concept code has been released. Refer to the GitHub Issue #7196 and Pull Request #7185 for technical context, and to the GitHub Commit 3749457 for the fix.

Detection Methods for CVE-2026-10800

Indicators of Compromise

Unexpected hash collisions or duplicate digest values produced by MultimodalHasher for inputs known to differ
Cache hits in FastDeploy multimodal pipelines that return outputs inconsistent with the supplied inputs
Local processes writing crafted multimodal payloads to FastDeploy input directories used by hash_features

Detection Strategies

Inventory deployed FastDeploy versions and flag any instance at or below 2.4.1 for review
Audit application code that consumes outputs of hash_features to determine whether the digest is used for security-relevant decisions
Compare hash values produced before and after applying commit 374945747652a8d32965591c0c01a00c88b7067f against a known test corpus

Monitoring Recommendations

Log invocations of MultimodalHasher.hash_features along with input metadata to support post-incident analysis
Alert on local user accounts writing to FastDeploy model or feature directories outside of expected automation paths
Track FastDeploy dependency versions in software bill of materials (SBOM) tooling to surface vulnerable installations

How to Mitigate CVE-2026-10800

Immediate Actions Required

Apply the upstream patch by updating FastDeploy to a build that includes commit 374945747652a8d32965591c0c01a00c88b7067f
Restrict local access to hosts running FastDeploy so that only trusted service accounts can submit multimodal inputs
Review downstream consumers of hash_features output and stop using the digest for any integrity or authentication decision until the patch is deployed

Patch Information

The maintainers fixed the issue in commit 374945747652a8d32965591c0c01a00c88b7067f of the PaddlePaddle FastDeploy repository. The corresponding pull request is tracked at GitHub Pull Request #7185, with discussion in GitHub Issue #7196. Additional metadata is published at VulDB CVE-2026-10800.

Workarounds

Limit FastDeploy execution to isolated, non-multi-tenant hosts to remove untrusted local actors from the threat model
Wrap or replace the hash_features output with a stronger digest such as SHA-256 in caller code until the patched release is adopted
Disable caching or deduplication paths that depend on MultimodalHasher output when handling untrusted multimodal inputs

bash

# Configuration example
pip install --upgrade "fastdeploy>2.4.1"
# Or pin to a commit that includes the fix
pip install "git+https://github.com/PaddlePaddle/FastDeploy.git@374945747652a8d32965591c0c01a00c88b7067f"