We tried to jailbreak our AI (and Model Armor stopped it)

Try out Model Armor → https://goo.gle/3KTnreG
Model Armor sample code → https://goo.gle/4nuaBB4

“AI Security” may sound intimidating, but it doesn’t have to be. What if it was just an API call? Martin Omander sits down with Aron Eidelman to demonstrate Google’s Model Armor. In this video, the duo try to jailbreak an app, make it leak sensitive data, and feed it malicious URLs. Watch how Model Armor blocks these threats before they even hit the model, and then see the actual code to implement it.

Chapters:
00:00 – Why AI apps need a "bodyguard"
00:57 – What are the top AI security risks? (OWASP Top 10)
01:46 – [Demo] Trying to jailbreak our AI app
02:25 – [Demo] Stopping sensitive data (SSN) leaks
03:23 – [Demo] Redacting data instead of blocking (DLP)
04:06 – [Demo] Blocking malicious URLs
04:50 – How it works: A simple API call
05:11 – Code: Sanitizing user prompts (Input check)
05:21 – Code: Sanitizing model responses (Output check)
06:19 – Code: Redact sensitive data
08:11 – Q&A: Don’t models already have guardrails?
07:23 – Q&A: Why not use another LLM to protect my LLM?
07:58 – Q&A: Configuring policies for different apps
08:50 – Q&A: How much does Model Armor cost?
09:10 – Final thoughts

Resources
Model Armor docs → https://goo.gle/3L0Me0s
Model Armor pricing → https://goo.gle/3IW3qUd
Aron’s and Martin’s video about sensitive data protection → https://goo.gle/3KYGeVP

Watch more Serverless Expeditions → https://goo.gle/ServerlessExpeditions
🔔 Subscribe to Google Cloud Tech → https://goo.gle/GoogleCloudTech

#GoogleCloud #Security #Serverless

Speakers: Martin Omander, Aron Eidelman
Products Mentioned: Model Armor