ailef

Hi HN!

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!

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?

He didn't imply it's morally dubious, I just read it as "people have more pressing matters to direct their attention to than this".

Notice the hypocrisy of the "explosions reported" title instead of "US bombs Venezuela".

Not paying rent is kind of a problem for most people.

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...

To me it seems out of step with US foreign policy.

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.

Isn't it "hustle"? But since it's a hassle to hustle it kinda makes sense too.

Prophetic bitcointalk thread from 2011:

https://bitcointalk.org/index.php?topic=53855.0

The cookies are for showing you targeted advertising. They are not related to the functioning of the website (except the necessary cookies).

filetype:txt should do the trick!

The assessment that the US is more likely to invade another country than china is not ridiculous anymore

When was it ever?

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...

https://news.ycombinator.com/item?id=33755016

How would an LLM agent "block these security bugs at runtime" exactly?

And your colleagues really spent a few days trying to figure this out?

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.

For the admin part, check out http://snapadmin.dev

It's my open source project which provides similar functionality, although not nearly as complete (and as well maintained) as the Django alternative.

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).

Yes, because with all the history of coups, invasion and illegal wars the US is definitely the side that wants "peace and prosperity"...

If there's something you can easily say on HN without the slightest piece of evidence and without fear of being called out is implicating Russia in any situation like this.

This thread reads like a propaganda fair sponsored by Western governments.

*in a perfect world with no monopolies, no information asymmetry and probably a dozen other caveats.

There's no need to define it as hate speech, since the DSA purposely mentions a very vague "harmful speech/content" (not "hateful") which can in turn applied to whatever the EU sees fits.

For example, any objection to the current policies w.r.t the Russian invasion of Ukraine can be censored due to the DSA and this has happened recently to a prominent Italian YouTube influencer (Andrea Lombardi). His tweets were censored for his opinions about the Ukraine war.

How many other tweets are being removed which we don't know of because they are not coming from famous people who have channels to make themselves heard anyway?

Reddit used to be really easy to scrape. I haven't really had the need to do it after the API changes drama, but the trick of appending `.json` to the URL apparently still works, example:

https://www.reddit.com/r/TheWire/comments/1aqvtoz/every_year...

Not sure if this only works on posts or also entire subreddits.

You write a text containing the JSON with all the data and instructions on how to analyze it and you instruct the LLM to extract what you're looking for (the "prompt"). The easiest way to start is probably the OpenAI API, or you can try one of the self-hosted LLM although I've never interacted with them programmatically.

These are all fair points I agree with. Let's just say I might've started writing code without thinking too deeply about the problem...

About the last point (slowness), this is definitely a problem right now. In my case, also retrieving the information from Amazon takes a sizeable amount of time; for the LLM part I think it can be optimized with finetuned models and performance will generally get better in the future. So maybe overall slowness is the easiest problem to solve.

You can build stuff just for fun...

They are just a bunch of floats, and they are usually stored in a binary non-human-readable format.

Part of the job especially when there are either tight deadlines or bad bugs

It's fine if this happens occasionally on deadlines or unexpected problems.

But it definitely shouldn't be part of any (good) job unless your contract explicitly provides for this possibility (with extra compensation).