TRUST // SECURITY

Security architecture

Threat model

Tamga scans LLM traffic in-line. Our threat model maps one-to-one onto the OWASP LLM Top 10 (2025): prompt injection (LLM01), sensitive information disclosure (LLM02), improper output handling (LLM05), excessive agency (LLM08), model denial of service (LLM04) and agent-tool attacks in chained models (LLM07).

Defence in depth

1. Input scan: PII/PCI/secret/injection scanners run on an Aho-Corasick DFA, targeting sub-5 ms.
2. Policy engine: YAML-based rule engine — versioned, two-person approved, rollback-capable.
3. Output scan: Canary tokens plus PII/secret scanners are applied to the response body too, with a sliding-window strategy for streaming responses.
4. Circuit breaker: Provider short-circuiting, rate-limit and budget-cap protection.

Cryptography

• TLS 1.2/1.3 (ECDHE + AES-GCM / ChaCha20-Poly1305).
• Optional client-cert auth (mTLS) — enabled for regulated customers.
• Audit log hash chain: SHA-256(prev_hash || canonical_json(entry)).

Incident response

T+0–1 h: detection and triage. T+4 h: containment and client notification. T+72 h: notification package prepared for the Turkish DPA under KVKK Article 12. T+30 days: post-mortem and control gap closure.