PaddleOCR/PaddleX Arbitrary Code Execution via unvalidated eval() in NormalizeImage

Vulnerability Overview Title: Arbitrary code execution via eval() of attacker-controlled scale string in NormalizeImage (text detection preprocessing) #5210 Description: Vulnerability Type: Arbitrary Code Execution Vulnerability Location: method in Root Cause: When loading a model directory, the parameter in the file is not validated and is executed directly via , allowing attackers to inject malicious code. Trigger Condition: The attacker can control the file in the model directory, specifically the parameter. Impact Scope Affected Products: - PaddleInference/PaddleX: v3.7.2 (latest release, 2020-06-25) and HEAD commit ff64096423708863ff508da312a5c0d2a7f462 - PaddleInference/PaddleOCR 3.x: Indirectly affected via the dependency Attack Scenarios: - Any channel that allows users to load model directories, such as third-party model errors, model sharing platforms, model serving pipelines, etc. - Attackers can execute arbitrary Python code by modifying the parameter in the file. Remediation Development Status: Not assigned, no label, no milestone, no project. Development Direction: Restrict text detection scale expressions. Participants: PaddleOCR/PaddleX POC Code Affected Products and Versions PaddleX up to v3.7.2 (latest) and HEAD commit ff64096423708863ff508da312a5c0d2a7f462 PaddleOCR 3.x affected via dependency >=3.7.0,<3.8.0 Additional Information Status: Open Labels: None Milestone: None Project: None Relationships: None Development: Restrict text detection scale expressions Participants: PaddleOCR/PaddleX Summary This vulnerability allows attackers to execute arbitrary Python code by controlling the parameter in the file within the model directory. Affected products include PaddleInference/PaddleX and PaddleOCR 3.x. The specific fix is not yet defined, but the development direction is to restrict text detection scale expressions.

Goal Reached Thanks to every supporter — we hit 100%!

PaddleOCR/PaddleX Arbitrary Code Execution via unvalidated eval() in NormalizeImage

More from this source