Hello World

Hello World

│by Charles
announcementsecurity

I’m Charles, the founder of farfa.tech. PAF has been my nights-and-weekends project for the past 6 months — not my full-time job, but something I deeply believe in. In the early days of AI, no one cared about security.

Now that AI has become part of organizations, security needs to be part of AI.

PAF — Prompt Application Firewall — is an attempt at reducing prompt injection and jailbreak by using AI models and other algorithmic techniques.

I believe most of the industry will have to buy a PAF in the upcoming years, similar to how the WAF became a standard on web APIs.

The product is in early access — not all the desired features are there yet to make a full product

BUT Thank you for trying it!

Pricing:

FAQ

Do you host and train your own models?

Yes, we train and host models at the moment. The challenge is that they must be able to detect dangerous prompts under a second — so they must be very fast and accurate.

They all run on dedicated Farfatech infrastructure in Canada, with encryption at rest and in transit.

How do you test PAF?

I have a dataset of 1,500 prompt injections and jailbreaks, plus 10,000+ regular prompts.

The goal is 0% false positive rate on regular prompts and 100% detection on dangerous ones. As of June 2026, PAF currently achieves: 9% false positive rate and 75% detection on dangerous prompts.

How do I integrate PAF into my application?

A simple API call before calling your LLM. An SDK is planned.

What happens when PAF detects a dangerous prompt?

The response will contain a match, indicating the prompt is dangerous.

Which AI providers/models does PAF support?

All providers. PAF integrates with your backend, so it’s provider-agnostic.

Is PAF itself vulnerable to prompt injection?

No. The architecture and tech stack make it impossible. This was my number one concern — simply wrapping a large context model on top of another means both can be prompt-injected.

Cheers,

Charles, Founder

back to blog