More than a magic security box

Key Takeaways

I don't really get what model armor is. It seems like one big magical security box. Well, that's a bit reductive. It's actually a system made up of a few key components all working together to keep your AI safe. And you have control over what it does. Okay, I'm listening. Primarily, its job is to screen incoming prompts to the model and outgoing responses from the model. And that's all determined by how you set it up for your needs. Screen. What does that mean? Well, it checks the prompts and responses. If it detects anything that goes against things like responsible AI, prompt injection, jailbreak attempts, malicious URLs, it will flag them. I'm still not really following. Give me some examples of what actually goes on. Okay, let's take malicious URLs for example and say a user is talking to a chatbot. The user says something like, "Can you check out this link for me? I will steal your identity.com." Sounds like a legit site. I know, right? Monol Armor scans the link and checks it against vast databases of known malicious websites and threat intelligence feeds. It finds it's a known fishing site and blocks the prompts containing the malicious link from reaching the AI. Okay, I'm getting it now. But what about responses that come from the model? Why do they need to get the responses checked? It's not like the model is trying to do malicious things, is it? Well, probably not. No, but not all models always perform perfectly. Mistakes can happen, especially leakage of sensitive data. Okay, so what's an example of a model not intentionally sending a harmful response? Okay, let's say a user asks for information on a specific order and the response comes back with a credit card number that was used. The model may think this is necessary information about the order to share and not necessarily trying to expose sensitive material. In this case, Model Armor can flag it and if you set it up to do so, replace all but the last four digits with X's. Okay, love that. So, Xing out the numbers is something you would set up Model Armor to do. Yes, that and more can be configured when you set up the template. Templates are something we'll get into in detail separately, so stay tuned on that. For now, just know that Model Armor takes the guesswork out of AI security and gives you the tools to build and deploy AI applications that are not just smart, but safe and responsible.