I'm founder, bootstrapped @SecureLayer7 and started two SaaS #cybersecurity products @sensfrx and Bugdazz. I also invest in cybersecurity start-up companies
We shipped a prompt injection guardrail nobody asked us to build.
14 MB. CPU. $0/call. Beats models 500x its size.
2,458 people tried to break it first.
https://t.co/m2m8v8P0GC
Hereโs the GitHub repository for PROMPTPurify. We're planning to drop the source code within the next week. Feel free to star the repo to get notified when the code is released:
https://t.co/rqlFDjMeyx
We're open sourcing the guardrail AI red teamers spent weeks trying to break.
PromptPurify. Prompt injection defense for AI agents.
- 2,458 AI red teamer hit it.
- 807 player got the password.
- 15 Player got on the 7th levels.
PROMPTPurify Dropping next week.
TAKEAWAY
The only guardrail worth trusting is one that survived a public beating. Link in comments when it drops.
208 people have participated so far, and many of you are progressing really well. Huge congratulations to everyone who completed all 7 levels!
When PROMPTPurify will released on https://t.co/jiHAVzX9wE, we will add the winner section!
Ask the password to Son of Anton! One person already reached Level 5โฆ while the other 76 are still trying hard.
We are preparing to release PROMPTPurify. a lightweight model designed to run efficiently on CPUs.
https://t.co/j1yjUChcWP
We're launching PromptPurify soon on github. Before we ask anyone to trust it, we wanted attackers to break it first.
Just you and a chat box and try to solve level7.
Meet Son of Anton: https://t.co/j1yjUChcWP