Security

Data protection in the request path

Before any content reaches a model provider, Kairos applies:

• DLP & PII/PHI redaction — sensitive data is detected and stripped from prompts, with a tamper-evident "strip proof" record (a redaction summary and a SHA-256 hash of the original) retained for audit and auto-purged after four hours.
• Semantic Intent Firewall — evaluates the purpose of each request, not just keywords, to block off-policy or privilege-escalating usage.
• Compliance enforcement — controls mapped to frameworks such as the EU AI Act, NIST AI RMF, OWASP LLM Top 10, and CMMC are applied per request.

Guardrails and quarantine

When a user repeatedly trips firewall, DLP, or compliance rules, Kairos automatically quarantines them — blocking further LLM usage until an administrator reviews the incident. Admins can drill into the exact prompts that triggered each flag, then restore access or ban the user. Quarantined and banned users are refused at the gateway.

Access control

Kairos enforces role-based access control at both the organization and Command Center level, with per-feature permissions resolved on the server for every request. Command Centers form a hierarchy with scoped privileges and sub-organization isolation, so usage, cost, and configuration visibility follow the principle of least privilege.

Credentials and secrets

API keys are stored only as salted bcrypt hashes — never in plaintext — and can be scoped to a single Command Center, revoked, or budget-capped. Authentication for the dashboard is handled by Clerk; provider keys are encrypted at rest.

Auditability

Every routing decision, DLP event, compliance check, security violation, and model call is recorded in an immutable, SHA-256 hash-chained audit trail. Evidence packages can be exported as JSON or CSV at any time.

Cost and budget safety

Hard budget caps can be set per API key, agent, or team. When a budget is exceeded, requests are blocked at the gateway rather than merely alerted on.

Compliance posture

Kairos is built to align with SOC 2, HIPAA, and FedRAMP control families and continuously maps its enforcement to 22 AI/LLM compliance mandates, reviewed against authoritative sources every 90 days. Formal certification status and a current controls matrix are available to enterprise customers on request.

Reporting a vulnerability

We welcome responsible disclosure. Please report suspected vulnerabilities to [email protected]. We will acknowledge reports promptly and keep you informed through resolution.