I built Before Upload, a small browser-based tool that checks files before you upload them, in general, or specifically to AI tools.
The tool tries to identify mostly PII leaks (personal info, VAT codes, API keys/secrets) and prompt injection attacks and runs locally in the browser. There is no backend analysis, no account, and the file is not uploaded anywhere.
Right now it supports:
- DOCX
- XLSX
- PPTX
- PDF
- JPG / PNG
- TXT / Markdown / HTML
It's still early and definitely not perfect. Right now, I'm building an internal dataset as I go, with several examples of prompt injection attacks hidden in PDFs and other Office file formats and I'm working to improve detection against it. I'm also looking at the next possible features like OCR, bulk analysis, and others, but it's still not clear what direction to take.
Known limitations:
- no OCR yet;
- PDF hidden text detection is incomplete;
- prompt injection detection is rule-based and can be noisy;
- it can produce false positives, especially on AI/security documents;
The current version is more of a warning tool than a cleaner. There are a few toggles in the advanced settings to customize the search behaviour and e.g. disable prompt injection detection on visible text because it can generate a lot of false positives, especially if the text is technical and naturally contains instructions/code.
I'd really like feedback to understand whether people could find such a tool useful and in which environments.
> Hitting companies with monetary fines does not work. Hitting the employees with jail time will make sure they don't sign on dangerous or known problematic systems.
What!? So, when you can't switch jobs because the market is bad or for any other reason, your choices are: 1) quit and lose the income (which you can't afford) or 2) sign on whatever and accept the risk of jail time?
> We take an opinionated view of the Spring platform and third-party libraries so you can get started with minimum fuss
Spring Boot is definitely opinionated (this is taken from their home page). Maybe not as much as RoR, but saying it isn't at all sounds very strange to me, having worked with it for a few years too...
The difference is you need a lot of training data to do that. Instead, now you can just tweak a system prompt and adapt it to whatever new policy you want to implement.
I've built something similar a few years ago, combining Wikipedia content and open domain pictures/videos to create long form videos automatically. Uploaded a bunch on YouTube as well. Wrote a blog post in case anyone is interested: http://ailef.tech/2020/04/29/turn-any-wikipedia-article-into...
Thanks! I think the Django one allows more stuff like e.g. layout customization. I haven't used it extensively so it was more of an assumption honestly! It's been around a long time and I just spent a couple months on my project instead.
What would be the advantage of using an LLM here? As opposed to parsing/validating the input using any other technique (of it's a well defined format, where you have e.g. a grammar).
I built Before Upload, a small browser-based tool that checks files before you upload them, in general, or specifically to AI tools.
The tool tries to identify mostly PII leaks (personal info, VAT codes, API keys/secrets) and prompt injection attacks and runs locally in the browser. There is no backend analysis, no account, and the file is not uploaded anywhere.
Right now it supports:
- DOCX
- XLSX
- PPTX
- PDF
- JPG / PNG
- TXT / Markdown / HTML
It's still early and definitely not perfect. Right now, I'm building an internal dataset as I go, with several examples of prompt injection attacks hidden in PDFs and other Office file formats and I'm working to improve detection against it. I'm also looking at the next possible features like OCR, bulk analysis, and others, but it's still not clear what direction to take.
Known limitations:
- no OCR yet;
- PDF hidden text detection is incomplete;
- prompt injection detection is rule-based and can be noisy;
- it can produce false positives, especially on AI/security documents;
The current version is more of a warning tool than a cleaner. There are a few toggles in the advanced settings to customize the search behaviour and e.g. disable prompt injection detection on visible text because it can generate a lot of false positives, especially if the text is technical and naturally contains instructions/code.
I'd really like feedback to understand whether people could find such a tool useful and in which environments.
Thanks!