Redact

Privacy redaction toolkit for images, PDFs, Word documents, and PowerPoint presentations. Use when the user needs to redact, mask, or replace sensitive/priva...

This package appears to be a legitimate local redaction/OCR toolkit. Before installing, note: 1) It requires heavy ML packages (paddlepaddle/paddleocr/paddlex) that may download large binaries or model files during install or first run and may require a lot of disk space and time. 2) The scripts create a cache directory (~/.cache/redact_temp) and temporary result directories; inspect and clean these if you need to avoid leaving extracted content on disk. 3) The code sets DISABLE_MODEL_SOURCE_CHECK in its environment—this is likely non-malicious but unusual; review the scripts if you have strict execution policies. 4) Run the tool in an isolated virtual environment (uv creates one) and inspect sample runs on non-sensitive files first. There are no declared network endpoints, no requested credentials, and no obvious exfiltration code in the bundle, but be aware that model downloads will require network access during install/use.

Type: OpenClaw Skill Name: redact Version: 0.1.1 The skill bundle provides legitimate privacy redaction tools but contains high-risk code patterns that are vulnerable to exploitation. Specifically, `read.py` and `redact-presentation.py` use `subprocess.run` to execute system-level commands via PowerShell (Windows), AppleScript (macOS), and LibreOffice for document conversion. While the scripts attempt some basic sanitization (e.g., escaping quotes), the construction of these command strings is fragile and potentially vulnerable to command injection if an attacker provides a file with a specially crafted name. No evidence of intentional malice, such as data exfiltration or hardcoded backdoors, was found; however, the reliance on shell execution for core functionality poses a significant security risk.