What it does
Built for CoreWeave Hacks, SIEGE uses TypeSafe to check the agent's proposed actions.
Playable demos & bots
Two hundred people try to break an agent's defenses while a defender uses each breach to improve them.
Only you can see your notes.
Screenshot unavailable. Open the experiment ↗
Built for CoreWeave Hacks, SIEGE uses TypeSafe to check the agent's proposed actions.
Start by writing clear rules for what your assistant can do. For example, list limits like maximum refund amounts. A developer can copy this project's setup to build a safety checkpoint. Before the assistant changes any records, it must check each requested action against your list of rules.
Your developer can run the project on their computer in test mode without paid accounts. Turning on live checks requires an access key that connects the app to TypeSafe. Remember that automated filters can still make mistakes, so blocked actions should offer a human helper.