Smart Guard

• 

  Smart Guard

Inspiration

We noticed prompt injection can compromise agent systems, so we set out to make them safer.

What it does

Smart Guard checks user input sent to AI, stopping harmful prompts before they cause problems.

How we built it

We use an embedding model with a classic LLM guard to catch risky prompts, balancing detection speed with accuracy.

Challenges we ran into

Finding the right trade-off between catching more attacks and keeping the system fast.

Accomplishments

Smart Guard blocks 92% of jailbreak attempts and costs three times less than other solutions.

What we learned

Combining different detection methods makes systems safer without slowing them down.

What's next

We will extend support, improve detection for new threats, and help more users secure their AI agents.

Built With

• html
• python