Zero-Trust by Design

MalikClaw is built on the principle that your AI agent shouldn't have unchecked power. Every action is cryptographically verified and locally contained.

Threat Model

MalikClaw is engineered to protect against:

  • ▹
    Prompt Injection AttacksStrict separation of instruction memory and user input contexts.
  • ▹
    Unauthorized ExfiltrationDefault-deny network policies ensure skills can only communicate with approved endpoints.
  • ▹
    Privilege EscalationAgent runs in unprivileged user space without root access requirements.

Cryptographic Verification

Signed Actions

Critical transactions are signed using local Ed25519 keys generated upon installation. The private key never leaves your device's secure enclave (or equivalent keystore).

Skill Integrity

Skills fetched from the registry are checksummed and signature-verified before execution. Tampered skills are automatically rejected.

Human-in-the-Loop (HITL)

High-risk actions require explicit human approval via the local TUI or web interface before the agent can proceed.

💳 StripePayments & Refunds
📧 GmailSending external emails
🏢 OdooModifying ERP records

Local-First Data Flow

Your data never leaves your device without explicit consent. Vector embeddings and memories are stored in your local workspace.

Local Device

Memory & Keys

LLM Provider

(e.g. Gemini/OpenAI)

Audit Logs

Every API call, file access, and state change can be recorded in an immutable local ledger for complete transparency.

malikclaw agent --audit-log

Vulnerability Reporting

Security is an ongoing process. If you discover a vulnerability, please report it securely through our dedicated channel.

Read SECURITY.md