Security Overview¶

AlexClaw is a single-user system designed to run on infrastructure you control. Security is layered — each component has independent defenses that don't rely on other layers being intact.

Attack Surface¶

Defense Layers¶

1. Authentication & 2FA — session-based web auth, TOTP verification for sensitive operations
2. Agent Authorization — PolicyEngine with capability tokens, chain depth, audit logging
3. Dynamic Skill Sandbox — namespace enforcement, permission model, process isolation, AST-based external detection
4. Encryption at Rest — AES-256-GCM for API keys and tokens in the database
5. MCP Authentication — Bearer token with constant-time comparison
6. MCP Policies — tool-level restriction rules for MCP clients
7. Content Sanitization — 7-layer heuristic pipeline strips prompt injection payloads from external content before LLM ingestion. Hidden HTML/CSS detection, zero-width unicode stripping, 101 known patterns from runtime JSON (Garak), imperative tone heuristic. See SECURITY.md

Key Principles¶

• No hardcoded secrets — all sensitive values come from environment or encrypted DB
• Constant-time comparison — all token/signature checks use Plug.Crypto.secure_compare/2
• Fail closed — missing config = feature disabled (not open)
• Audit everything — authorization decisions are logged with caller, permission, and outcome
• Let it crash — OTP supervision handles failures; security checks don't swallow errors