Never-Leak Protocol Specification
The complete technical specification for the Never-Leak Protocol. Nine chapters defining how AI agents interact with secrets without those secrets ever entering the agent's context window.
Overview
Goals, architecture, terminology, design principles, and conformance tiers.
Agent Identity
Level 1Cryptographic identity, Agent Identity Documents (AID), attestation, trust levels (L0-L3), and lifecycle management.
Action-Based Access
Level 2Typed action semantics, opaque handles, scope grants, conditions, and secret rotation propagation.
Execution Isolation
Level 3Isolated subprocesses, environment scrubbing, output sanitization, memory wipe, and timeout enforcement.
Pre-Execution Defense
Level 4Command interception, injection detection, exfiltration pattern matching, and policy validation.
Audit Integrity
Level 5SHA-256 hash-chained records, HMAC signatures, tamper-evident verification, and forensic queries.
Attack Detection & Response
Level 6Behavioral anomaly detection, prompt injection detection, circuit breakers, and automated response.
Cross-Agent Trust & Federation
Level 7Delegation tokens, scope attenuation, trust federation, and result-only propagation.
Wire Protocol & Transport
JSON message format, transport bindings (HTTP, stdio, WebSocket), MCP integration, and error handling.
Conformance Tiers
Agent identity, action-based access, execution isolation.
Adds pre-execution defense and immutable audit trails.
Full protocol with attack detection and cross-agent trust.