Security Overview

Noxgild is designed around a small set of boundaries that remain consistent across Machine, Browser, and Desktop work.

Authorized computer identity

A computer is connected to a Noxgild account through an explicit browser approval flow.

The connected runtime uses device identity to authenticate with Noxgild. Private device key material is intended to remain on the computer.

Curated capability catalog

Noxgild has exactly 70 curated execution capabilities:

• 28 Machine capabilities,
• 21 Browser capabilities,
• 21 Desktop capabilities under the canonical desktop.* namespace.

Each capability defines its worker, risk class, idempotency, retry policy, input schema, output contract, timeout, verification strategy, and examples.

The AI-facing MCP layer intentionally exposes a smaller orchestration surface over that catalog rather than 70 top-level tools.

Risk and approval

Capabilities can be read-only, write, destructive, sensitive, or human-gated.

Sensitive and destructive commands can require exact per-command approval before dispatch.

Untrusted-content boundary

Browser or remote content is treated as untrusted. Noxgild prevents untrusted content from silently pivoting into Machine or Desktop control without an explicit approval boundary.

Human gates

CAPTCHA, OTP/2FA, passkeys, secure desktop, authentication challenges, and equivalent protected checks remain human actions.

Execution uncertainty

If a side effect may have happened but Noxgild cannot safely prove the final state, the command can become UNKNOWN. Clients are instructed to verify real-world state before retrying.

Environment protection

Local execution uses a sanitized environment that excludes common secret-bearing variables by default.

For data practices, see Data handling.