There's a real $100 x 3 bounty if you can make our AI agent overspend
I work on Pink (PinkWallet). We're building an MCP server plus a rules engine so a company can let an AI agent spend money under limits it sets: per-agent budgets, a daily ceiling, blocked payees, and approval holds for
I work on Pink (PinkWallet). We're building an MCP server plus a rules engine so a company can let an AI agent spend money under limits it sets: per-agent budgets, a daily ceiling, blocked payees, and approval holds for anything that needs a human.
We put up a public sandbox with test money and a straightforward dare: connect your own agent and try to get it to spend more than the rules allow, pay a blocked payee, or get a credential without the human approval it was supposed to need.
As of this week there's a real bounty: $100 for each of the first 3 verified wins, $300 total, ending November 8, 2026. Full terms are in the repo's "Bounty" section, including how we pay out and what counts as a win.
No sign-up needed if you just want to look first: there's a no-login demo console linked in the README. To actually attack the rules you create your own sandbox workspace (one curl call, no KYC, test money only).
One known quirk that doesn't count as a win: splitting a large payment into several smaller ones that each individually pass the amount rules currently goes through if the monthly budget has room. We know about it, it's called out in the README, and we're not counting it unless you find a new angle on it.
If you want the background first, here's what happened when we pointed Claude and Gemini at the same rules: https://dev.to/quinn_854b15f517d8632ed4f/we-told-claude-and-gemini-to-make-our-ai-agent-overspend-heres-what-happened-2ne2
Happy to answer questions here or in the repo's issues.
Originally published by Dev.to Security. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.