Skip to content

Security: harlixay7/AgentXFlow

Security

SECURITY.md

Security Policy — AgentXFlow

AgentXFlow is local-first developer infrastructure designed to coordinate multiple autonomous coding agents on the same Git repository safely.


1. Threat Model

AgentXFlow operates as an authoritative local control plane binding strictly to 127.0.0.1 (loopback only).

Security Boundaries

  • Loopback Enforcement: The Model Context Protocol (MCP) daemon binds strictly to 127.0.0.1 by default and validates Host and Origin headers to prevent unauthorized cross-origin requests from web browsers.
  • Dynamic Per-Install Token: On first launch, AgentXFlow generates a cryptographically secure 256-bit authentication token saved in local data storage (.agentxflow/auth.token). No hardcoded tokens exist in source code or production builds.
  • Zero Self-Certification: Autonomous agents cannot mark their own criteria valid or bypass verification. MCP requests to self-satisfy criteria (criteria_satisfy) are rejected.
  • Isolated Git Worktrees: Agents operate strictly inside dedicated Git worktrees located at .agentxflow/worktrees/task-<id> or in the coordinator AppData worktree pool. Agents never directly mutate the active working tree or main branch.
  • Attempt-Scoped Mutation & Scope Auditing: On task submission, the coordinator checks git diff --name-only and git status --porcelain against granted exclusive file locks (scope_leases) for the active attempt. Any unreserved file modification triggers a scope violation rejection.
  • Server-Controlled Machine Verification: An agent cannot mark a task complete by assertion. The coordinator executes verification profiles and machine evaluators (cargo test, npm test, compiler checks) under its own process supervision and computes a deterministic SHA-256 evidence digest.

2. Token Management & Rotation

  • The active authentication token can be inspected in the UI or retrieved via the Tauri IPC get_mcp_info command.
  • Users can rotate the active token at any time via the UI or rotate_mcp_token command.
  • When configuring agents (OpenCode, Claude Code, Antigravity, Codex, Cursor), supply the Bearer token in the Authorization header:
    Authorization: Bearer <your_token>

3. Reporting a Vulnerability

If you discover a security vulnerability or bypass in AgentXFlow:

  1. Please do not open a public GitHub issue.
  2. Email security reports directly to the maintainer: harlixay7.
  3. Include a description of the vulnerability, reproduction steps or adversarial test case, and affected versions.
  4. We will acknowledge receipt within 48 hours and work with you on a coordinated disclosure timeline.

There aren't any published security advisories