The Php Times

Intercept, created by Victor Ukam for the Laravel AI SDK, introduces a middleware pattern for AI prompts. Positioned between the agent and the AI provider, it allows developers to inspect, rewrite, or reject prompts before they leave the application. It is designed as an additional security layer rather than a replacement for existing input validation or access controls.

The package currently provides two primary guardrails. The PromptInjectionGuard uses regular expressions to detect manipulation attempts, such as instructions to ignore previous prompts. Developers can choose to block the request, log the event, warn the user, or sanitize the content. Advanced usage allows for custom callbacks to log detections or prepend specific instructions to the prompt.

To protect user privacy, the PIIRedactor identifies structured personal data, including emails, phone numbers, IP addresses, API keys, bearer tokens, and credit cards (via Luhn check). Depending on the configuration, it can mask, redact, log, or block the prompt. By default, highly sensitive data like credit cards and API keys trigger an immediate block unless configured otherwise.

Intercept requires PHP 8.4 and the laravel/ai package. While it currently ships at version 0.1.4, the roadmap includes future expansions into observability, performance, and guidance. Configuration can be handled per agent via the HasMiddleware interface or globally through a published config file.