|
| 1 | +# PAD-003: The "Identity Sidecar" Pattern & JIT Intent Signing for AI Agents |
| 2 | + |
| 3 | +**Publication Date:** January 03, 2026 |
| 4 | +**Author:** Vouch Protocol Maintainers |
| 5 | +**Subject:** Decoupled Cryptographic Signing for Stochastic AI Models |
| 6 | +**Status:** Public Prior Art |
| 7 | +**License:** Apache 2.0 |
| 8 | + |
| 9 | +## 1. Abstract |
| 10 | + |
| 11 | +This disclosure places into the public domain a security architecture for Large Language Models (LLMs) known as the "Identity Sidecar." This pattern solves the critical security risk of entrusting non-deterministic, hallucinatory models with long-lived private keys. Instead, the cryptographic identity is isolated in a deterministic "Sidecar" process (e.g., an MCP Server) that performs Just-In-Time (JIT) signing only when explicitly requested and validated against a local policy. |
| 12 | + |
| 13 | +## 2. Problem Description |
| 14 | + |
| 15 | +In standard agentic architectures, developers often inject API keys or private keys directly into the LLM's system prompt or environment variables. This creates two failure modes: |
| 16 | + |
| 17 | +1. **Key Leakage:** The LLM may accidentally output the private key in its response (Prompt Injection). |
| 18 | +2. **Unauthorized Usage:** If the LLM enters a loop or is jailbroken, it can use the keys to perform unlimited actions without checks. |
| 19 | + |
| 20 | +## 3. The Solution: The Identity Sidecar Pattern |
| 21 | + |
| 22 | +We disclose a method where the "Agent" is composed of two distinct processes: |
| 23 | + |
| 24 | +1. **The Brain (Stochastic):** The LLM (e.g., Claude, GPT-4) which reasons and plans. It holds **ZERO** cryptographic secrets. |
| 25 | +2. **The Passport (Deterministic):** A local sidecar service (e.g., Vouch MCP Server) that holds the `Ed25519` private keys in secure memory. |
| 26 | + |
| 27 | +### Architecture Diagram |
| 28 | + |
| 29 | + |
| 30 | + |
| 31 | +### 3.1 The "Just-In-Time" (JIT) Signing Flow |
| 32 | + |
| 33 | +The signing process is inverted. The Agent does not "login" at the start. Instead: |
| 34 | + |
| 35 | +1. **Reasoning:** The LLM decides it needs to perform an action (e.g., "Book Flight"). |
| 36 | +2. **Request:** The LLM sends a structured request to the Sidecar: *"Please sign this specific intent payload: {'action': 'book', 'amount': 500}."* |
| 37 | +3. **Policy Check (The Guardrail):** The Sidecar evaluates the payload against deterministic logic (e.g., "Is amount < $1000?"). |
| 38 | +4. **Signing:** Only if the policy passes, the Sidecar signs the payload and returns the signature to the LLM. |
| 39 | +5. **Execution:** The LLM attaches the signature to its API request. |
| 40 | + |
| 41 | +### 3.2 Security Properties |
| 42 | + |
| 43 | +| Property | Traditional Approach | Identity Sidecar | |
| 44 | +|----------|---------------------|------------------| |
| 45 | +| Key exposure to LLM | Direct access | Zero access | |
| 46 | +| Prompt injection risk | Critical | Mitigated | |
| 47 | +| Rate limiting | Application-level | Cryptographic | |
| 48 | +| Audit trail | Logs only | Signed intents | |
| 49 | + |
| 50 | +## 4. Application to Model Context Protocol (MCP) |
| 51 | + |
| 52 | +We specifically disclose the implementation of this pattern via the **Model Context Protocol (MCP)**. |
| 53 | + |
| 54 | +* The **Vouch MCP Server** acts as the Identity Sidecar. |
| 55 | +* The **MCP Client** (IDE/Chatbot) connects to this server. |
| 56 | +* The LLM uses the `vouch_sign` tool exposed by the MCP server to obtain cryptographic proofs on demand. |
| 57 | + |
| 58 | +### 4.1 Example MCP Tool Definition |
| 59 | + |
| 60 | +```json |
| 61 | +{ |
| 62 | + "name": "vouch_sign", |
| 63 | + "description": "Sign an intent payload with the agent's cryptographic identity", |
| 64 | + "inputSchema": { |
| 65 | + "type": "object", |
| 66 | + "properties": { |
| 67 | + "intent": { |
| 68 | + "type": "object", |
| 69 | + "description": "The action and parameters to sign" |
| 70 | + } |
| 71 | + }, |
| 72 | + "required": ["intent"] |
| 73 | + } |
| 74 | +} |
| 75 | +``` |
| 76 | + |
| 77 | +### 4.2 Example Usage |
| 78 | + |
| 79 | +```python |
| 80 | +# The LLM requests a signature through the MCP tool |
| 81 | +result = await mcp_client.call_tool( |
| 82 | + "vouch_sign", |
| 83 | + {"intent": {"action": "book_flight", "amount": 450}} |
| 84 | +) |
| 85 | + |
| 86 | +# The sidecar returns a signed token (or error if policy fails) |
| 87 | +vouch_token = result.content[0].text |
| 88 | + |
| 89 | +# LLM uses the token in API call |
| 90 | +response = await api_client.book_flight( |
| 91 | + headers={"Vouch-Token": vouch_token} |
| 92 | +) |
| 93 | +``` |
| 94 | + |
| 95 | +## 5. Prior Art Statement |
| 96 | + |
| 97 | +This architecture combines "Hardware Security Module (HSM)" principles with "Agent Tool Use." By applying the HSM pattern specifically to the **context window limitations and non-deterministic nature of LLMs**, we establish this as a distinct architectural standard for the AI Agent era. |
| 98 | + |
| 99 | +### 5.1 Related Work |
| 100 | + |
| 101 | +| Concept | Source | Differentiation | |
| 102 | +|---------|--------|-----------------| |
| 103 | +| HSM/TPM | Industry standard | Applies pattern to LLM context | |
| 104 | +| Sidecar pattern | Kubernetes/Service Mesh | Applies to AI agent identity | |
| 105 | +| MCP Tools | Anthropic | Novel use for cryptographic signing | |
| 106 | + |
| 107 | +### 5.2 Claims Established as Prior Art |
| 108 | + |
| 109 | +This disclosure precludes patents on: |
| 110 | + |
| 111 | +1. "LLM Key Isolation" - Separating cryptographic keys from LLM context |
| 112 | +2. "Proxy-based Agent Identity" - Using a proxy service for agent authentication |
| 113 | +3. "JIT Signing for AI Agents" - On-demand cryptographic signing for agent actions |
| 114 | +4. "Policy-gated Agent Signing" - Applying deterministic policy checks before signing |
| 115 | + |
| 116 | +## 6. Reference Implementation |
| 117 | + |
| 118 | +- **Repository:** https://github.com/vouch-protocol/vouch |
| 119 | +- **MCP Integration:** `vouch/integrations/mcp/` |
| 120 | +- **Documentation:** https://github.com/vouch-protocol/vouch#mcp-integration |
| 121 | + |
| 122 | +--- |
| 123 | + |
| 124 | +*This document is published as prior art to prevent patent assertion on the described concepts while allowing free use by the community under the Apache 2.0 license.* |
0 commit comments