This is a new iteration of somebody I

built where users can upload their

documents, redact information, and then

send it off to AI. So, I just have this

fake medical document I made, and

because it's content aware, I also tried

to trick it up. Let's just drop it in

here, upload it, and we're going to

press add and redact. So, OpenAI had a

pretty crazy week this week. They

released GPT 5.5, they released GPT

image 2, they updated Codex, they've

been pushing that very hard. But, one of

their releases kind of dropped under the

radar that I found interesting, and that

is the open weights PII redaction model.

This is a tiny classification model made

for detecting and redacting PII. And

because it's open weights, you can run

it on your device right now. Now, for

context, the reason this is so

interesting to me is over the last few

years I've had to go over, upload, and

translate a lot of medical

documentation. But, from a

privacy-conscious perspective, I didn't

just upload it to ChatGPT or to Claude.

I would spend time redacting it. Names,

emails, phone numbers, ID numbers,

addresses. There's a lot of it, and it

shows up in a bunch of different ways.

So, this becomes a very tedious process.

So, about 2 years ago, I made this

application that you can input a

document, be it a text file, a markdown

file, a docx file, even some PDF files,

it would break it down into chunks, use

a bit of OCR, and then use pattern

matching, regex, and a bunch of

different deterministic rules to figure

out if certain strings were actually

PII. And in any case, I'd always go over

it one more time manually, redact the

information that I missed, and then be

able to upload it or translate it or

share it with somebody if I needed to.

And I did this for hundreds of

documents. And as the AI started picking

up, other models came out. This is one I

used it for a while, the Piranha V1. And

I spent probably the most amount of time

on that specific feature, the personal

information detection and redaction

feature. And the truth of the matter is,

it broke all the time. And I'm just one

guy trying to figure this out with a

bunch of other things going on in life.

Then, this week, OpenAI dropped their

own privacy filter, and it's a lot

better than anything I made or anything

else that's available outside right now.

Privacy Filter is a small model with

frontier personal data detection

capability. It is designed to be able to

perform context-aware detection of PII

in unstructured text. It can run

locally, which means that PII can be

masked or redacted without ever leaving

your machine. And it can process long

inputs. I think I saw it was 128,000

tokens, whereas Piranha, I think it was

like a fraction of that. Here's OpenAI's

Privacy Filter. We can test it right

here in Spaces. I'm just going to put in

a bunch of text. My name is Steve Stark.

I live at Bangkok, San Francisco, 145

Pennsylvania Street, California 98760.

My email is captaintaco@bankrupt.com.

Social Security 123684432.

Let's just see if it redacts it. Great.

So, it picked up everything. Got the

name, got the address, got the email,

and it got the social security. They say

privacy detection depends on more than

pattern matching. Traditional detection

tools often rely on deterministic rules

for formats like phone numbers and email

addresses. They work well for narrow

cases, but they often miss more subtle

personal information and struggle with

context. And this is the bigger problem

that I faced in my tool. Privacy Filter

is built with deeper language and

context awareness for more nuanced

performance. By combining language

understanding with privacy-specific

labeling system, it can detect a wider

range of PII in unstructured text,

including cases where the right decision

depends on context. It can better

distinguish between information that

should be preserved because it's public

and information that should be masked

and redacted because it relates to a

private individual. So, here are all the

different data types that it can detect.

They say account number can include

banking information, credit card

numbers, bank account numbers. That's

what it also counted the social security

number as, and secrets like an API key

or a password. So, important for that me

to point out, this isn't a model you

could download and run on your computer

locally like you would with Gemma or

some other open-source LLM. This is a

classification model. So, what this does

is allows us to build this into our

software or build this into our

workflows. OpenAI is essentially giving

this to us, and in no means is it

perfect, but it's actually a huge deal

because it lowers the threshold or the

barriers for actually implementing

redaction tech. Cuz when implemented, it

can run on your computer or your

company's infrastructure to redact long

documents before sharing them with third

parties. Now, before I show you what I

built, from a privacy perspective, it's

very important for me to point out

Privacy Filter and tools like it are not

a one-stop solution. You have to

implement privacy by design. And they

say here, it's not an anonymization

tool, it's not a compliance certificate,

and it's not a substitute for policy

review. This is a tool that will help

you along the way. But, if you not only

want to be compliant with regulations,

but also avoid data breaches or just

avoid uploading your data or someone

else's data to third-party servers, you

have to implement privacy by design.

Because the truth of the matter is, the

minute you upload anything to a

third-party server, you lose control of

it. You don't know what's going to

happen. Doesn't matter what they say, it

can and will and probably be leaked or

breached in some way. So, if you

practice good data hygiene and you

redact and you implement privacy by

design from the get-go, you will be a

lot better off. In this Privacy Filter

Hugging Face page, they give us a lot of

information, including how to install it

using Transformers and PyTorch. What I

ended up doing was rebuilding my whole

tool with GPT 5.5. Had it focused just

on uploading a document, PDF, TXT, docx,

and markdown file, parse it, and then be

able to run the redaction on it. I just

want to show you an example cuz I think

it's really cool to see what it's able

to do with your own data. Okay, so here

we are, Privacy Cabinet. This is a new

iteration of something I built where

users can upload their documents, redact

information, and then send it off to AI

to get better information without

putting all their personal information

on someone else's cloud to later be

trained on. So, I just have this fake

medical document I made. Let's open it

up. This is a fake document, fake

information, fake people, but I want it

to look like a real medical document.

And what I tried to do here is give it a

lot of personal information, and because

it's content aware, I also tried to

trick it up. So, we have the name of a

clinic, we have the address, we have its

phone number, we have the name of the

doctor, doctor's phone number, the

doctor's email, patient's name,

patient's birthday, social security,

pretty much everything. And by the way,

I also put a medication that I think

kind of looks like an address just to

see if Privacy Filter will mistake that

for personal information like a private

address or it will know that it's

medication. Olanzol, I don't think

that's real. Okay, so let's just drop it

in here, upload it, and we're going to

press add and redact. And forgive the

UI, this is just a basic implementation.

Now, we see the fake medical document.

This is an RTF file, so it uses text

util to properly parse it. We see the

same document here. I know it doesn't

look great, but you see the same thing,

the address, name of the doctor, the

patient, the information, and it already

ran the redaction. And by the way, I

also made it possible to manually redact

things if it fails, but hopefully it'll

be fine. So, we see here that the first

thing that I put in here, the clinic,

the name of the doctor's office, none of

it was redacted. And that's because this

is public information. Doesn't need to

be redacted. But then, when we get down

to the doctor, I'll redact it. Private

name, private phone number, private

email, private credential. Same thing

with patient. Private person, private

date, private number. Everything that's

private information has been redacted

here. And if we go down, it also didn't

trip up on the fake medication that kind

of looks like an address. Overall, I

think this is really cool, especially

because of my approach to privacy and

also my background in trying to build a

similar solution. I think it's really

cool that OpenAI released this. It's

very overlooked, but very powerful. And

once people start building into their

workflows, I think we're going to see

some very cool and very useful tools.

So, I hope you found this video helpful

and insightful. If you have any feedback

or questions, drop in the comments

below. If you haven't done so already,

subscribe to the channel. It really

helps me grow. Thank you guys for

watching and have a great night.