An agent's API key got stolen. In one minute, 11 million tokens were gone.
Another team had two agents pinging each other in a loop — by the time anyone noticed, $1,800 had vanished.
Here's the real problem: every single request looked "normal," so no alarm ever fired.
Why didn't anything go off?
It's simple, actually. Every request an agent sends looks fine on its own — within budget, valid token, an approved merchant. A human trying 300 payments an hour trips fraud detection instantly. An agent doing the same thing at that speed just looks like a "system glitch," not fraud.
Gartner expects 1 in 4 enterprise breaches by 2028 to trace back to AI agent exploitation. And yet most companies only have dashboards and alerts in place. A dashboard is an observation tool, not an enforcement tool. The moment spending crosses a limit, something needs to block it — a dashboard just shows you the damage after the fact.
Card networks jumped into this fight too
So payment infrastructure moved first. In June 2026, Visa plugged its entire network into ChatGPT, letting agents complete purchases at any Visa merchant. Mastercard is pushing Agent Pay, built on an extension of tokenization called Agentic Tokens that only lets verified agents transact. Both stick to the existing four-party card model, keeping fraud protection and dispute resolution intact.
Coinbase took a completely different route. x402 revives the old HTTP 402 status code to let agents pay directly on-chain in stablecoin — no account, no card. It processed 169 million transactions in its first year alone. Then there's Google's AP2, layered on top of either rail — it cryptographically proves "this person actually authorized this transaction," regardless of payment method. Over 60 companies, including Adobe, Amex, Mastercard, and PayPal, are already on board.
| Card networks (Visa, Mastercard) | Stablecoin (x402) | |
|---|---|---|
| Best for | High-value consumer purchases | Tiny, high-frequency API payments |
| Dispute resolution | Existing card system | Still immature |
| 2026 scale | Visa's stablecoin pilot: $7B annualized | Coinbase: 169M transactions in year one |
So this is what showed up
While the card networks fight over standards, what teams actually need first is simpler: a layer that enforces exactly how much an agent can spend, where, and with whose approval — before the transaction happens.
Solo developer Pratik Pandey's Orite goes right at that gap. The tagline says it all — "Give your AI agent money. Not a blank check." It enforces budgets, policies, and approval requirements before money moves, logs every transaction to an audit trail, and expands an agent's autonomy step by step as trust builds. It's free to start, with a 3-month promo running right now.
Worth flagging though — it's early. The questions in the Product Hunt comments hit exactly the soft spot: what happens when a purchase is wrong but still within budget? How do refunds and disputes get handled? The direction is right; dispute resolution just isn't fully there yet.
And in practice, credentials get breached before cards do. That 11-million-token incident above was a stolen API key, not a stolen card. Which is why the table below matters as much as card locks.
| Shared card | Agent-only virtual card | |
|---|---|---|
| Blast radius if stolen | Spreads to other agents/uses | Contained to one card |
| Merchant restrictions | Basically impossible | Locked per transaction |
| Usage tracking | Unclear who spent what | Auto-attributed per agent |
Locking a single-use virtual card to a merchant and amount, and issuing separate credentials per agent, cuts the blast radius dramatically on its own.
Check these five things right now
You don't need to wait for the standards war to settle. If your agents already have payment or API access, start here.
- Ban shared cards
Issue a dedicated virtual card or credential per agent. If one gets breached, the rest stay safe. - Lock merchant and amount
Set a merchant whitelist and cap per transaction. General agents: $100–300/day. Sensitive tasks: under $10. - Tier your approvals
Auto-approve small amounts, flag mid-size ones, and require human sign-off for large amounts or new merchants. - Merge token costs with card spend
Put API token bills and card transactions on one dashboard, attributed by team and agent. Runaway costs usually blow up in tokens before they hit cards. - Review your agent registry quarterly
Document each agent's owner, purpose, budget, and approved merchants — and revisit it regularly.
Want to go deeper
Orite A trust layer that enforces budgets, approvals, and audit trails on agent payments orite.tech
6 Guardrails to Limit AI Agent Spending Six practical guardrails for agent spending on payment rails fystack.io
How to Set Spending Controls for AI Agents Ramp's practical checklist for virtual cards, approvals, and attribution ramp.com
Agentic Commerce in 2026: AP2, x402 and AI Payments A breakdown of how Google's AP2 and Coinbase's x402 differ structurally bitontree.com
Visa, Mastercard And Coinbase Are Fighting Over How AI Agents Pay An analysis of why card networks and stablecoin camps are colliding forbes.com
$1.8k before anyone noticed A real-world breakdown of how runaway agent spend slips through without triggering alarms requesty.ai
