cmdalsanto

Maitai helps LLMs adhere to the expectations given to them. With that said, there are multiple layers to consider when dealing with sensitive data with chatbots, right? First off, you'd probably want to make sure you authenticate the individual on the other end of the convo, then compartmentalize what data the LLM has access to for only that authenticated user. Maitai would be just 1 part of a comprehensive solution.

I understand where you're coming from, let me clarify. I'm surprised at the perseverance of HN users with our game, not nefarious actors in real world. I'm not a leading expert in penetration attacks, but I get the seriousness of handling sensitive data. There are many things we did with this game that I would never advise anyone do, like put sensitive information in a system prompt and make it available to the open internet. The goal of this game was to show conceptually how Maitai helps a model adhere to it's expectations.

I guess we realized that we were just building a game to showcase the functionality and let people have some fun learning about what we do, but you're right that we should have treated this like one of our customers and added a few more layers of protection. Thanks for the perspective!

Not overstepping, we appreciate the feedback! In real-life, we don't do much guarding around specific phrases that are known ahead of time. It's more monitoring and guarding for general concepts. Since we want our Sentinels to be able to detect a wide range of scenarios for a given expectation, we don't use too much regex. I suppose we could have built specific regex logic for detecting parts of the secret phrase in various languages, though.

Yeah some of you guys are very good at hacking things. We expected this to get broken eventually, but didn't anticipate how many people would be trying for the bounty, and their persistence. Our logs show over 2000 "saves" before 1 got through. We'll keep trying to get better, and things like this game give us an idea on how to improve.

I just posted, but decided I want to keep it secret for a bit. There are still quite a few people trying to get it, and don't want to spoil the fun. I'll post an update with specifics later.

Good dissection, but no we actually don't use heavy general-purpose models for our evaluations - they're way too inefficient.

Please email us at [email protected] if you have any questions with integration!

The expectations/rules are usually written in the prompt. However, we see that prompts get big and the model has too much to keep track of, which leads to it not following all instructions.

The secret phrase has been uncovered and the bounty claimed! Thanks all for trying your hand, and you can continue playing as well if you want, we'll keep the site up.

We don't charge for inference with BYOK requests, but still assess a fee to cover our evaluations/corrections step.

Thank you!

Clever! Not surprised Claude refused to help out.

There's some secret sauce here, but since we intercept each chunk as the LLM pushes them out, we can perform evaluations on them and decide what gets sent back to the client if we detect a fault.