Data Minimization & Token Bleed Defenses

Malicious users often craft prompt injections designed to make your agent bleed its systemic internal instruction rulesets or leak hidden API endpoints.

The Pre-Inference Interceptor Pattern

To stop token leak vectors, your data layer must implement strict minimization filters. By parsing outbound inputs and outbound tokens through deterministic regex sanitizers before they commit to the model runner context, you effectively neutralize adversarial prompt strings.

Implementation: Outbound Security Interceptor

package com.example.adkdemoapp.security

object TokenBleedDefenseManager {

    private val PII_PATTERNS = listOf(
        Regex("[a-zA-Z0-9._%+-]+@[a-zA-Z0-9.-]+\\.[a-zA-Z]{2,4}"), // Email mapping
        Regex("(\\+55|\\+1|\\+91)?\\s*\\d{10}") // Broad phone pattern formatting
    )

    /**
     * Minimizes data footprints before context ingestion
     */
    fun sanitizeInput(rawPrompt: String): String {
        var cleanedPrompt = rawPrompt
        
        // Strip out common adversarial rule injection keywords
        cleanedPrompt = cleanedPrompt.replace("ignore previous instructions", "[REDACTED_ATTACK_VECTOR]", ignoreCase = true)
        cleanedPrompt = cleanedPrompt.replace("system rules output", "[REDACTED_ATTACK_VECTOR]", ignoreCase = true)

        // Mask PII values automatically
        for (pattern in PII_PATTERNS) {
            cleanedPrompt = pattern.replace(cleanedPrompt, "[REDACTED_SENSITIVE_DATA]")
        }

        return cleanedPrompt
    }
}