Security and Privacy

Pheron Agent is built with a Privacy by Design and Sovereignty First approach. Because it operates with full system automation privileges, security checks are enforced at every step of execution.

1. On-Device Local Isolation

Pheron Agent is designed to run all planning, memory lookup, and execution loops on-device using local MLX models. When a local model is loaded and active, no prompt data or execution context leaves your machine.

• Local-First Default: All inference defaults to the on-device Titan Engine. A cloud fallback (OpenRouter) is only triggered when no local model is available. To enforce strict local-only operation, enable Local-Only Mode in Settings.
• Zero Telemetry: Pheron Agent has no tracking endpoints. Telemetry logs are stored locally at ~/Library/Application Support/PheronAgent/Telemetry/.
• Your files stay local: Screenshots, clipboard contents, and execution logs are never uploaded.

2. Privacy Guard

For workflows that request cloud integrations or use external fallback models, Pheron Agent routes requests through a security interceptor called Privacy Guard:

• PII Detection: A rule-based on-device scanner inspects prompt inputs for Personally Identifiable Information (such as names, addresses, credit cards, and system paths) before any external routing occurs.
• Enforcement Levels:
  • PASS: The prompt contains no sensitive info and continues.
  • DESENSITIZE: Automatically masks or replaces PII with placeholder strings (e.g. replacing actual names with [USER_NAME]).
  • BLOCK: Completely rejects the operation if high-risk data cannot be safely redacted, returning a safety warning.
• Performance Cache: Scans are cached securely using SHA-256 hashes with strict Time-To-Live (TTL) limits.

3. SecuritySentinel and GuardAgent

• SecuritySentinel: Enforces biometric verification (Touch ID / Apple Watch) before executing high-risk system commands or accessing the vault.plist secure Keychain configuration.
• GuardAgent: Evaluates generated code and shell outputs in real-time. If it detects unsafe patterns (such as unbounded recursive deletes or unauthorized network payloads), it immediately terminates the execution thread.

4. Privilege Sandboxing

To execute macOS system tasks (such as terminal scripts and system controls), Pheron Agent uses Apple's XPC Inter-Process Communication:

• The main user interface runs inside a restricted user environment.
• Privilege-requiring tasks are delegated to PheronAgentXPC, a notarized background helper service compiled with Apple's Hardened Runtime and strict entitlements.
• All actions are logged and subject to user accessibility approvals.