Most people giving AI agents crypto wallets are ignoring the obvious: what stops an agent from getting drained by a bad endpoint or a malicious prompt? If an agent sees a link and tries to pay it, right now it just signs. That’s a disaster waiting to happen. I spent the hackathon building Sober to fix this for @PayBox and @MoonPay. Think of it as a pre-flight safety check before your agent touches your money. When an agent decides to pay for an API or service, Sober steps in first: - It checks if the URL is legit HTTPS and not some raw IP or localhost drain script - It pings the endpoint to see if there is an actual valid x402 payment challenge behind it - It verifies the exact price and asset so the agent doesn’t overpay or approve the wrong token - Only if everything passes clean does it generate a PayBox signing request for your wallet In the video, you can see it in action: First, I try to get the agent to send money to a sketchy local drain link. Sober catches it instantly and blocks it. Then, I ask it to pay for a real GitHub lookup via glim.sh on Base USDC. Sober verifies the service, confirms the 0.002 USDC fee, and only then creates the live PayBox approval request. Zero blind signing. Your agent stays safe even if it runs into shady links. Check out the code on GitHub: github.com/Habuskid/sober