Security & privacy

A security product must be secure itself.

Threeum Defense is designed around least privilege, deny-by-default authority, tenant isolation, and local-first privacy. This page describes those controls and the current limitations of the alpha.

Authority model

  • Deterministic policy is authoritative. AI model output is advisory and cannot override policy or mint its own permissions.
  • Deny-by-default. Privileged agent actions require a fresh, short-lived, signed capability lease tied to a specific action, resource, and budget.
  • Bounded containment. Current containment acts only on a local registry (revocation) and a simulated sandbox adapter — no host firewall or process manipulation.
  • No self-promotion. Adaptive changes pass regression gates and human review before they take effect.

Data protection

  • Tenant isolation across data, threat graph, and keys.
  • Least privilege with roles, key rotation, and revocation.
  • Secrets stay out of logs and out of source control.
  • Bounded inputs — normalized, size-limited events with structured errors.
  • Audit trail of decisions, actions, and reversible registry mutations.

Privacy modes

Process locally; share only what's necessary

Local only

All processing stays on the device or in your environment. No raw data leaves.

Anonymized correlation

Share privacy-safe signals and derived metadata for cross-node correlation — not raw content.

Private cloud

Run in your own tenant with your keys and retention policy.

Air-gapped

Fully offline. Local protection continues without connectivity.

The Edge tier is designed to prefer derived metadata over raw content wherever that is sufficient.

The dashboard & your keys

The operational dashboard is an operator tool, not a consumer login. It holds your API key in memory only — never in localStorage, cookies, or the URL — and talks to a same-origin /v1 backend that you run.

  • The public static copy of this site cannot reach your localhost; it detects that and shows deployment docs instead of transmitting a key.
  • Before any authenticated call, it performs a /health capability handshake to confirm it is talking to a Threeum Defense service.
  • Untrusted event strings are rendered as text, never as HTML — no injection from event content.

Current limitations

  • All models are experimental research artifacts; published metrics are synthetic for six of the seven core families and public-dataset holdout for Edge and the TextGuard companion — not a guarantee of real-world efficacy. Swarm and Infrastructure exceed their false-positive budget at the frozen test threshold.
  • No third-party security certifications are claimed.
  • No production deployments or customers are claimed.
  • Containment is local and simulated only in this alpha — no OS or network enforcement.
  • Full production integrations, enterprise SSO, and high availability are in development.
Responsible disclosure. This is pre-release software. Security issues should be reported to the project maintainers through the repository's security policy once published.