The PCI Security Standards Council has released new guidance requiring human oversight for AI agents handling cardholder data to mitigate security risks and prevent unauthorized system access.
Key Points
- The PCI SSC recommends that human operators provide explicit approval for any AI actions involving cleartext cardholder data.
- Organizations should implement a "least agency" approach, limiting AI system permissions to only the specific capabilities required for assigned tasks.
- The guidance advises maintaining an AI inventory and bill of materials to track models, data-use policies, and unauthorized "shadow AI" tools.
- Security teams should conduct adversarial testing to identify potential vulnerabilities and ensure AI systems cannot bypass established access controls.
- Sensitive credentials, such as cryptographic keys and passwords, must be managed through independent tools rather than being stored in AI prompts or logs.
- Third-party AI providers must be assessed for compliance, with contracts explicitly prohibiting the use of company data for model training.