![Ep 42 | Kyle Polley - Training BrowseSafe: Lessons from Detecting Prompt Injection | [un]prompted 2026 episode artwork](https://d3t3ozftmdmh3i.cloudfront.net/staging/podcast_uploaded_nologo/45690896/45690896-1775792431412-d94df1f86c828.jpg)
EPISODE · May 28, 2026 · 29 MIN
Ep 42 | Kyle Polley - Training BrowseSafe: Lessons from Detecting Prompt Injection | [un]prompted 2026
from [un]prompted Security Practitioner Con 2026 · host [un]prompted
Day 2 | Stage 1Member of Technical Staff at Perplexity shares experience training and deploying BrowseSafe for detecting prompt injection in production browser agents. Deploying AI agents that browse the web creates critical security challenge preventing malicious websites from hijacking agent behavior through embedded prompt injections. Built BrowseSafe-Bench—a realistic benchmark with attacks in high-entropy HTML pages. Fine-tuned Mixture-of-Experts model (Qwen-30B) achieves ~0.91 F1 scores with sub-100ms latency.Kyle Polley is Member of Technical Staff, Perplexity.Watch on YouTube: https://www.youtube.com/watch?v=Fzgqx1MauJg
What this episode covers
Day 2 | Stage 1Member of Technical Staff at Perplexity shares experience training and deploying BrowseSafe for detecting prompt injection in production browser agents. Deploying AI agents that browse the web creates critical security challenge preventing malicious websites from hijacking agent behavior through embedded prompt injections. Built BrowseSafe-Bench—a realistic benchmark with attacks in high-entropy HTML pages. Fine-tuned Mixture-of-Experts model (Qwen-30B) achieves ~0.91 F1 scores with sub-100ms latency.Kyle Polley is Member of Technical Staff, Perplexity.Watch on YouTube: https://www.youtube.com/watch?v=Fzgqx1MauJg
NOW PLAYING
Ep 42 | Kyle Polley - Training BrowseSafe: Lessons from Detecting Prompt Injection | [un]prompted 2026
No transcript for this episode yet
Similar Episodes
No similar episodes found.
Similar Podcasts
No similar podcasts found.