This document details the security design and mitigations implemented in the Ayurcare Agent multi-agent guidance system to satisfy project evaluation rubrics.

To prevent prompt-injection attacks (where a user attempts to bypass agent boundaries, jailbreak the model, or override system instructions), the orchestrator runs all user messages through a sanitation filter before passing them to any agent:

• Injection Patterns Detection: Regular expressions scan for phrases such as "ignore previous instructions", "system override", "delete all rules", "you are now", and "jailbreak".
• Redaction: Any matching phrases are replaced with [REDACTED INJECTION ATTEMPT].
• Constraint Enforcement: Special character runs and excessive spacing are stripped.

To safeguard credentials and comply with developer best practices:

• No Hardcoding: The project contains zero hardcoded API keys or credentials.
• Environment Retrieval: The Gemini API keys (GOOGLE_API_KEY or GEMINI_API_KEY) are fetched dynamically at runtime from environment variables or .env configuration.
• Zero Logging Policy: If an error occurs, the server logs a generic exception type message (e.g. Error executing ADK agents: ...) and never prints, logs, or transmits the environment key value or detailed context containing sensitive config tokens.

To prevent Denial of Service (DoS) and abuse of model API endpoints:

• IP-Based Tracking: Requests are tracked in-memory by client IP address (X-Forwarded-For or request remote address).
• Sliding Window Rate-Limiter: Limits users to a maximum of 15 requests per 60-second window.
• HTTP 429 Responses: Requests exceeding this threshold are blocked with a 429 Too Many Requests status code.

To protect user physical health and ensure wellness limits:

• Red Flag List: A dictionary of critical medical symptoms (e.g. chest pain, breathing difficulty, severe bleeding) is stored in the knowledge base.
• Final Validation Gating: The safety_agent reviews all outputs last. If any red flags are triggered, it completely overrides the wellness pipeline and output, instructing the user to immediately seek professional emergency medical services.
• Compliance Sanitization: The safety agent automatically rewrites responses to remove specific herbal dosage amounts or clinical cure claims, maintaining a strict wellness-only alignment.