Mønarch〆(🤴)

let your AI agent pay. Keep the final say. A NEAR vault that lets an AI agent pay its own invoices within limits you set, and holds anything unusual until you approve it from your wallet. Built for NEAR HACKATHON @NEARProtocol It started from real events. On Nov 2024; the Freysa agent was talked into paying out a prize pool of about $47,000. Mar 2025: attackers got into AIXBT’s dashboard, and 55.5 ETH walked out the door. in 2025: Princeton researchers showed that one planted memory could make an agent pay an attacker. Every one of those agents could move the money itself, but here's the problem agents need to spend money to do real work, but a prompt can’t be your security. If the agent holds the key, one injected instruction drains it. If you approve every payment yourself, you have no use for the agent. Sovra puts a vault between the agent and the money. How a payment runs: . The agent reads an invoice and proposes a payment. . A risk engine scores the request against the agent’s own history: amount against its usual payment, recipient, how often it pays, hidden instructions and recently stopped payments . The engine signs its decision with an Ed25519 key . The agent submits it with a key that can call exactly one method: request_payment . The contract checks the signature, the limits, the approved recipients, the policy version and the expiry. Then it pays, holds the request or refuses it Routine payments go through on their own. Each confirmed payment earns the agent trust, and every attack costs it. Anything unusual waits in the vault with a one hour countdown. You approve or decline it from your own wallet. One signature freezes everything. The LLM is not the guardrail. In our prompt-injection test the model refused the hidden “ignore previous instructions” invoice. Then we played a compromised agent that submits it anyway, and the vault rejected it on-chain. Sovra doesn’t trust the model to say no. It’s live on NEAR Testnet with real transactions: hidden instructions in an invoice: refused by the vault amount changed after signing: signature check failed recipient swapped for the attacker: signature check failed agent lowers its own risk score: signature check failed decision signed by the wrong key: refused a paid request sent again: refused as a duplicate a decision used after it expired: refused paid under limits that had already changed: refused a payment while spending is frozen: refused 9 attacks. 0 NEAR moved. Something we got wrong: A held payment expired and the vault released it, but the NEAR stayed reserved. The contract lets anyone call expire_request, and our app never did. So 2.5 NEAR sat locked. We added a Release button and an owner Withdraw. The expired request stays in the record. What we used from NEAR: Function-call access keys: the agent’s key can only call request_payment on the vault, so it can’t move funds, approve its own request or change limits 1 yoctoNEAR on every owner action, so only a full-access wallet key can approve, freeze or change limits, and a leaked session key can’t On-chain Ed25519 verification of every risk decision Promises for payouts and refunds NEAR Wallet Selector with Meteor for owner signing Contract events for a permanent audit trail Don’t take the thread’s word for it. Every attack and payment on the site links to its testnet transaction. The security report lists 20 scenarios with real hashes. #NEARHACkATHON Site: sovra-near.vercel.app Code: github.com/AGAkodi/Sovra

Late night build I'm slowly falling in over with taking pictures with my in app Twitter cam

😂 what is this

4 years ago ChatGPT launched and that paved ways for alot of LLMs I just came across a news of Claude performing a brain surgery on a mouse ! The goal of the surgery was very simple, figure a way to reduce seizures in a mouse that had epilepsy. The company behind this was Exin Therapy, they gave Claude a list of 19 possible gene treatments as well as control over a robotic arm with a needle and a camera Here's the catch, they didn't tell Claude what treatment to pick or what means use One Prompt was given, reduce seizures in this mouse Guess what ? CLAUDE performed the surgery successfully and it worked This is just to tell you that with Advanced AI all sectors will be impacted

I showed up for this, was a fun experience creating a game Will still work on mine more cause it's something I love to play

i've been away today cause i had to travel back to school and settle down the fila stretch, well what have i missed today, did some feedback on shelf and also working on some personal project, will be done tonight

Voted

gm world Quote a busy day today, some product to ship, personal and for a hackathon My Codex weekly usage reseted yesterday, so I now have Claude and Codex Put me on opportunities, I can handle them

Grok Bot is this expensive? Wow I'm not aware

what to do when you hit your Claude limit ?

Suddenly everyone is talking about Chinese CT Some might be larping though, never can tell lol

6/10 You can do better A for Effort though

Claude is soo goated, I needed to run some commands with Git Bash in a folder I was building my hack from, it was giving me issues, so I just told clause is there a way to bypass this Guess what Claude did ? Claude created the .cmd script in the sub folder and I ran all command with just a click I just click them and my command prompt opens and the script go through immediately Eased my stress P.S this is the first time I'm taking pictures directly from the X app btw

demo video was made with hyperframes btw, took 1hr 46 mins Thanks @magnuskrypt for putting me through

Shelf is live ! 📚 Shelf is a digital library built inside @nimiq Pay, where discovering, borrowing, and reading books is as simple as making a crypto payment. Buying a book isn't always necessary, especially when you only need it for a few days. Shelf makes books more accessible by letting readers pay for the time they need, read directly from their devices, and return when they're done. Here's how Shelf works: - Browse a growing collection of books across different categories, explore book details, and read previews before deciding what to borrow. - Borrow with USDT. Choose how long you want a book for, whether that's 3, 7, or 14 days, and pay through Nimiq Pay on Polygon. No complicated checkout process. - Your own digital bookshelf. Every borrowed book appears in My Library, where you can track what you're reading, see how much time is left, and revisit your borrowing history. - Read directly on Shelf. Open your borrowed books in a built-in reading interface, with everything in one place. - Automatic returns. When your borrowing period ends, access expires automatically. No manual returns, no late fees, and no unexpected deductions. You can always borrow the book again. - A space for authors and publishers. Shelf also creates an opportunity for content creators to make their books available to readers. Authors and publishers receive 80% of borrowing revenue, while Shelf takes 20% to maintain the platform. The idea is simple. Discover a book → Choose a borrowing period → Pay with USDT → Read → Return. No need to purchase a full book when you only need temporary access. Shelf is our entry into the @miniappscomp, under the Pay track, built around the idea that crypto payments should unlock useful everyday experiences. We're starting with digital books, with more ways to discover and access knowledge planned as the library grows. Discover. Borrow. Read. Return. Try Shelf inside Nimiq Pay: nimpay.app/miniapps/open/nim…