Transparency
Threat Model & Attack Mitigation
How Spineforge protects your credentials from both insider threats and external attacks.
1. Internal Threats (Compromised Agents)
- Credential Harvesting: An agent cannot dump the Vault. It never receives raw provider keys at all: the Gateway injects them server-side, and only for the providers and models the agent's scopes allow.
- Blast Radius: If a gateway key (
sfgw-) is leaked, you can revoke it instantly from the dashboard. The attacker only loses access to that single agent's identity; they do not gain access to the raw OpenAI/Google keys. - Budget Limits: Agents are bound by strict cost ceilings. A hijacked agent cannot rack up a massive bill before being blocked by the Gateway's real-time cost tracker.
2. External Attacks (Against Spineforge)
- DDoS & Abuse: The Gateway and Registry APIs sit behind an aggressive Web Application Firewall (WAF) and global rate limiters. Malicious bursts are blocked at the edge before hitting our application servers.
- Vault Encryption: Your raw LLM keys (OpenAI, Anthropic, etc.) are encrypted at rest using AES-256-GCM. They are only decrypted in memory during the split-second when the Gateway proxies a request.
- Tenant Isolation: All telemetry data and configuration is logically isolated by Organization ID via Row-Level Security (RLS) in our PostgreSQL databases, preventing cross-tenant data leakage.