Threat Model

This document outlines the potential attack vectors against AI agent credentials and explains how Ephos mitigates these risks through its architecture.

Core Assumptions

  • Users follow security best practices for their Master Passphrase.
  • Clerk's authentication infrastructure is trusted for identity verification.
  • Cloudflare's runtime isolation is trusted for volatile memory handling.

Attack Vectors & Mitigations

VECTOR POTENTIAL IMPACT EPHOS MITIGATION
Prompt Injection Agent exfiltrates raw .env secrets to an attacker. Agent only has scoped Ephos Tokens; raw keys never enter the agent's context.
Server Compromise Database leak of all stored API keys. Zero-knowledge storage; encrypted blobs are unreadable without user passphrases.
Token Theft Stolen token used to drain API balance or exfiltrate data. Domain whitelisting restricts token use to authorized hostnames only.
Runtime Memory Dump Decrypted keys scraped from server RAM. Secrets exist in RAM for milliseconds; isolation via Cloudflare Workers prevents cross-request leaks.
Echo-Attack Third-party API leaks the key back in a response body. Automated response scrubbing redacts any sensitive patterns before return.

Trust Boundaries

Ephos explicitly defines where trust begins and ends:

  • User-Side: You trust your browser and your passphrase memory. Ephos never touches the unencrypted secret here.
  • Transit: You trust TLS encryption. Ephos only transmits encrypted blobs.
  • Execution: You trust the Ephos Enclave to handle the secret in volatile memory and wipe it immediately.