With the development of AI agents, machines are increasingly capable of controlling the mouse, manipulating the browser and performing many online tasks instead of humans. Giving AI the power to make payments is therefore becoming the next step.

But there is a big barrier: it is difficult for users to feel secure in directly providing credit card numbers, CVV security codes or passwords to AI. This data is at risk of being saved on the technology provider's server or stolen through command injection attacks.

When AI pays without seeing the card number

To solve this problem, the Bot Vault security mechanism was implemented on the Grok Bot autonomous agent engine version 0.75.2 October 28, with an approach that completely separates the AI ​​model from sensitive payment data.

According to Tuoi Tre Online's research, instead of letting AI directly read card information, the system stores the credit card number (PAN), expiration date, CVV code and authentication information in an independent security area (Vault). Data is encrypted and protected right on the device or in a trusted execution environment.

Meanwhile, the large language model (LLM) that drives AI in the cloud cannot see this information. When asked to find and book airline tickets or make online purchases, AI can still open the browser, select products, and fill in passenger information. However, at the payment step, instead of entering the actual card number, AI only uses a reference code (token).

When the user submits the payment form, Bot Vault's intermediary layer in the new browser intervenes, retrieves card information from the encrypted data store and automatically fills in the payment page. Thanks to that, the transaction is still performed without the AI ​​model directly accessing the card number or security code. This is a way to limit the risk of financial data being exposed through AI processing requests.

Secure the card number, but who keeps the wallet?

The "blind safe" mechanism completely solves the problem of exposing card data to hackers, but cybersecurity experts immediately warned of a new operational flaw: The card is protected safely, but money in the account can evaporate because of AI's "enthusiasm".

To ensure that the automated shopping process is not interrupted, Bot Vault integration tools allow for continuous execution mode, meaning that the AI ​​is allowed to call the payment order itself without needing the owner to re-approve it in the final step. The risk lies in the machine's ability to read the interface and context.

In reality, if the user issues a command that lacks strict limits, or the sales website installs dark patterns that cause the AI ​​to mistakenly identify the "Buy with premium service package" button as the "Continue" button, the system will still automatically open the "safe" for payment.

More dangerously, if the website has a slow network response error causing the AI ​​to think the transaction has failed and automatically execute a retry loop, it can silently click to pay again dozens of times for the same item in just a few tens of seconds. When all operations take place automatically at computer speed, users only discover the problem when their phone is bombarded with messages to deduct money from the bank.

Before autonomous AI tools complete the "emergency brake" mechanism at the hardware level, such as requiring biometric authentication (fingerprint, face) of the device owner for each card decryption or locking session spending limits, users who want to experiment with leaving shopping to AI should only link with virtual cards or prepaid cards with a small amount of money pre-loaded for each transaction.

Maximum:

1500 characters There are currently no comments. Be the first to comment