Project 02 — Infrastructure
Kybernos
Zero-trust gateway for AI agents. Every tool call authenticated, authorized, validated, and audited before it runs.
A drop-in security layer for agentic AI. NIST-aligned zero-trust architecture that intercepts every LLM tool call and forces it through an 8-service mesh of independent security controls. Works with Anthropic, OpenAI, Bedrock, Vertex, or any local model.
The problem
LLMs can now read files, run queries, and call APIs. That power is also the attack surface—prompt injection is OWASP's #1 LLM risk two years running. Most "AI firewalls" respond with another opaque AI you're asked to trust blindly. Teams need a security layer they can actually defend, not another black box.
What Kybernos does
It applies NIST SP 800-207 Zero-Trust Architecture directly to AI tool-calling. The model is never the security boundary. Every tool call clears an independent chain: authenticate → authorize → validate → sandbox → audit. RBAC and deterministic schema validation are authoritative. Period.
What you get
-
1
NIST-aligned zero-trust pipeline
Eight single-purpose services across four trust zones. One compromise doesn't break the whole chain.
-
2
Per-call capability tokens
Fresh, scoped, short-lived ES256-signed tokens, revocable in real time. Access is per-request, not standing.
-
3
Authoritative schema validation
Path traversal, non-SELECT SQL, and non-HTTPS URLs are structurally impossible. No model can inject them.
-
4
112-rule semantic firewall + 2,867-probe verified
Every block attributed to the exact rule. Zero false-positives on benign traffic. Proven in adversarial corpus.
-
5
Encrypted, tamper-evident audit
AES-256-GCM audit trail. Full forensic chain of custody for every request, every response.
-
6
Model-agnostic routing
Native Anthropic adapter. OpenAI-compatible edge for OpenAI, Ollama, vLLM, LiteLLM, anything. Switch providers without touching security controls.