SECURITY & GOVERNANCE POLICY

Mirrabase Policy Framework

Detailed documentation of our security controls, data handling rules, PII redaction standards, and vulnerability disclosure policies.

POLICY 01 / DATA SECURITY & REDACTION

Inline PII & Secret Redaction Policy

Mirrabase inspects incoming prompt payloads and outbound model responses in memory before transmission. Sensitive credentials—including API tokens, passwords, credit card numbers, and Social Security Numbers—are redacted using high-speed regular expression heuristics and Named Entity Recognition (NER).

Redacted tokens are replaced with standardized placeholders (e.g., [REDACTED_API_KEY]) before the payload reaches external LLM provider endpoints, preventing accidental exposure of sensitive customer data in third-party model training logs.

POLICY 02 / THREAT SHIELDING

Prompt Injection & Adversarial Filtering

To protect autonomous AI agents from malicious system prompt overrides and jailbreaks, Mirrabase applies multi-layer threat filtering. System instructions are enforced at the proxy gateway level and isolated from user input fields.

Requests containing known adversarial attack signatures are logged and rejected with a standardized HTTP 400 Bad Request error before reaching model inference endpoints.

POLICY 03 / ACCESS CONTROL

Virtual API Keys & Budget Limits

Engineering teams issue virtual API keys mapped to specific applications or environments. Each virtual key can be configured with strict rate limits (requests per minute) and token budget caps.

When a virtual key exceeds its designated monthly budget, Mirrabase automatically throttles or revokes traffic for that key while keeping other production services unaffected.

POLICY 04 / VULNERABILITY DISCLOSURE

Responsible Vulnerability Disclosure

We welcome security researchers and developers to inspect our open-source codebase. If you discover a security vulnerability in Mirrabase or Tuenel Gateway, please report it directly to security@mirrabase.ai. We commit to acknowledging reports within 24 hours and providing patch timeline updates.