ThirdKey provides an enterprise‑grade trust infrastructure that enables security teams to safely deploy autonomous AI agents. Its stack includes a Rust‑native, policy‑governed runtime with Cedar authorization and tamper‑evident audit logs, cryptographic schema verification (SchemaPin) anchored to DNS, and domain‑anchored agent identities (AgentPin) that issue short‑lived credentials for capability‑scoped access. Customers can adopt individual primitives or the full suite to protect sensitive data while using AI agents in production.
Funding
Funding not disclosed
Founders
Product
Problem
Enterprises adopting autonomous AI agents face security risks such as unauthorized tool usage, supply‑chain attacks, and lack of enforceable runtime policies, making it difficult for security teams to approve and monitor agent deployments.
Solution
ThirdKey delivers an enterprise‑grade trust infrastructure that lets security teams safely deploy autonomous agents. Its Rust‑native runtime, Symbiont, enforces policy‑governed execution using Cedar authorization, sandboxed environments, and tamper‑evident audit logs. SchemaPin provides cryptographic verification of agent and server schemas to block malicious “MCP rug pull” attacks, while AgentPin anchors agent identities to a domain with short‑lived ES256 credentials and capability‑scoped access. ToolClad adds declarative contracts for tool interfaces, defining typed parameters, validation, and policy in a single manifest. Together, these primitives create a verifiable, auditable supply chain that CISO‑approved agents can operate against sensitive data.
Target Audience
Primary customers are enterprise security teams, CISO offices, and AI platform operators that need to enforce policy, verify agent integrity, and maintain auditability for autonomous AI deployments.
Features
- Symbiont: Rust‑native runtime with Cedar policy enforcement, sandboxed execution, and immutable audit logging
- SchemaPin: ECD5‑P256 signed schema verification with `.well-known` DNS anchoring and TOFU key pinning
- AgentPin: Domain‑anchored cryptographic identity, short‑lived ES256 credentials, and 12‑step verification protocol
- ToolClad: Declarative `.clad.toml` contracts specifying typed I/O, validation, output parsing, and Cedar policies for CLI, session, and browser tools
- Interoperable, self‑contained primitives that can be adopted individually or as a full stack