Before Upload started from a simple observation:
people are uploading more and more files into AI systems without really knowing what those files contain.
A PDF can hide metadata, annotations, or failed redactions. A Word document can contain comments and tracked changes. An Excel spreadsheet can have hidden sheets or formulas. A PowerPoint deck can include speaker notes. Even screenshots and images may expose EXIF metadata, GPS coordinates, account IDs, or sensitive information in the background.
And now, in the AI era, files can also contain hidden instructions designed to influence how an AI system behaves.
Before Upload exists to make this problem visible.
The project focuses on one specific moment:
the few seconds before a file is uploaded to an AI tool.
The goal is simple:
help users understand what an AI system might read inside a file before sharing it with ChatGPT, Claude, Gemini, Grok, Copilot, Perplexity, or any other AI assistant.
This blog is where we research and document:
- hidden data inside common file formats;
- prompt injection attacks and indirect prompt injection;
- real-world metadata and redaction failures;
- AI security incidents;
- PDF, DOCX, XLSX, PPTX and image privacy risks;
- AI agent vulnerabilities;
- safe workflows for AI-assisted document handling;
- practical file inspection techniques.
The focus is practical rather than theoretical.
Most articles are written for developers, security researchers, AI builders, privacy-conscious users, consultants, and anyone working with documents and AI systems in real workflows.
Before Upload itself is built around a few principles:
- local-first processing;
- no unnecessary file uploads;
- transparency over magic;
- simple explanations;
- practical security;
- privacy by default.
The project is still evolving.
Current limitations exist, false positives happen, and some file formats are harder to inspect reliably than others. The goal is not to promise perfect security. The goal is to make hidden file layers easier to notice before they become a problem.
The core idea behind everything on this site is simple:
If a machine can read it, you may be sharing it.Before Upload started from a simple observation:
people are uploading more and more files into AI systems without really knowing what those files contain.
A PDF can hide metadata, annotations, or failed redactions. A Word document can contain comments and tracked changes. An Excel spreadsheet can have hidden sheets or formulas. A PowerPoint deck can include speaker notes. Even screenshots and images may expose EXIF metadata, GPS coordinates, account IDs, or sensitive information in the background.
And now, in the AI era, files can also contain hidden instructions designed to influence how an AI system behaves.
Before Upload exists to make this problem visible.
The project focuses on one specific moment:
the few seconds before a file is uploaded to an AI tool.
The goal is simple:
help users understand what an AI system might read inside a file before sharing it with ChatGPT, Claude, Gemini, Grok, Copilot, Perplexity, or any other AI assistant.
This blog is where we research and document:
- hidden data inside common file formats;
- prompt injection attacks and indirect prompt injection;
- real-world metadata and redaction failures;
- AI security incidents;
- PDF, DOCX, XLSX, PPTX and image privacy risks;
- AI agent vulnerabilities;
- safe workflows for AI-assisted document handling;
- practical file inspection techniques.
The focus is practical rather than theoretical.
Most articles are written for developers, security researchers, AI builders, privacy-conscious users, consultants, and anyone working with documents and AI systems in real workflows.
Before Upload itself is built around a few principles:
- local-first processing;
- no unnecessary file uploads;
- transparency over magic;
- simple explanations;
- practical security;
- privacy by default.
The project is still evolving.
Current limitations exist, false positives happen, and some file formats are harder to inspect reliably than others. The goal is not to promise perfect security. The goal is to make hidden file layers easier to notice before they become a problem.
The core idea behind everything on this site is simple:
If a machine can read it, you may be sharing it.