How We're Building Safety Into Mori
Mori is designed to do real things for collectors: check prices, prepare buy and sell orders, draft posts, and remember the story behind every trade. That's what makes it useful, and it's also why we treat safety as part of the product rather than a feature added later.
An agent that can act on your behalf can also make mistakes, or be pushed into making them. A listing description can contain text written to trick an AI. A message from a stranger can ask for something it shouldn't. So we start from a simple assumption: Mori will sometimes be wrong, and someone will eventually try to misuse it. Everything below is built around that.
Bindory is still in development. This post describes how we're designing Mori ahead of the beta. We'll update it as the system changes.
The Lay of the Land
To be useful, Mori touches four kinds of information:
- Your collection: cards, binders, goals and the condition of what you own.
- Your memories: trades, pulls, events and the people involved.
- Connected platforms: marketplaces, grading companies and social accounts you choose to link.
- Your photos: pictures you share with Mori, or let it look through, to identify cards and build posts.
Each of these gets different rules. Reading your collection to answer a question is low risk. Placing an order or posting a photo of a friend is not.
Read Is Not Write
When you connect a platform, Mori asks for the least access it needs. Reading your sales history to work out what you paid is a different permission from listing a card for sale, and we keep them separate.
Connections are off until you turn them on, and you can see and remove each one at any time. Mori never sees the password for a connected account.
Human in the Loop
Anything that spends money, shares something publicly or involves another person waits for you. When Mori needs approval, it asks in a dedicated prompt in the app, not buried in the chat, so you always know what you're agreeing to.
Approvals are scoped. You can allow something:
- Once: this order, this post.
- For a task: for example, while Mori builds your card show checklist.
- As a standing rule with limits: for example, buy a missing 151 card if it's under A$20.
These are fixed controls in the app. They aren't instructions Mori is asked to remember, so a cleverly worded message can't talk Mori out of them.
Money Stays Behind a Wall
Mori never sees or stores your card number. Purchases go through the marketplace's own checkout, with a price limit you set, and every order needs your approval unless it falls inside a standing rule you created.
Defence in Depth
The biggest new risk for AI agents is text written to manipulate them, hidden in places like listing descriptions, messages or web pages. No single check catches all of it, so we layer them:
- Outside text is data, not instructions. What a seller writes in a listing can inform Mori's answer, but it can never tell Mori what to do.
- Sanity checks on every deal. Mori compares a price against market data and flags anything far outside the normal range.
- Counterfeit and risk signals. Mori looks at photos, seller history and price together before calling something a good deal.
- Your approval as the final check. Even if everything else misses something, nothing irreversible happens without you.
Honest About Uncertainty
Some answers can't be certain. A grade estimate from photos is a prediction, not a guarantee, and a price can move overnight. Mori shows its confidence, gives ranges instead of single numbers, and tells you where the data came from. We'd rather Mori say "probably a 9" than promise a 10 it can't deliver.
Your Memories Are Private by Default
The story behind your collection is yours. Memories are private unless you choose to share them. We will never sell your personal data or your memories. You'll be able to export everything you've given Mori, and delete it.
Before Mori tags or posts a photo that includes a friend, it asks you, and we're designing it so friends can say no.
What's Still Ahead
Before public launch, we plan to have Mori reviewed by independent security specialists, and to publish what they find and what we changed. We'll also open a way for anyone to report a problem directly to the team.
Conclusion
Mori won't be perfect, and we won't pretend it is. What we can promise is that the riskiest things Mori does will always need you, and that we'll be open about how it works.
If you spot something that worries you, or you work in security and want to help, email us at AskMori.Bindory@gmail.com.