AJ NiPlexAdarsh Jaiswal

Home / Projects / NiPlex-Harness / Safety model

Pictures

Safety model

Safety is several small checks, not one wall. They reduce mistakes and block known bad patterns. They are not a security boundary against a determined attacker.

Since 2 Sept 2026 · Updated 29 Oct 2026

Risky things wait for a button

Layers

  1. Allow-list: who may talk to the bot.
  2. Approval gate: risky code and commands wait for a button.
  3. Hard blocks: some actions never run.
  4. File protection: file tools cannot touch core code.
  5. Sandbox limits: time, memory, processes.
  6. Secret vault: placeholders and output redaction.
  7. Content screening: injection-like text is flagged.
  8. Page reader guard: non-public addresses are refused.

Honest limit

Checks match patterns. New phrasing or a clever script can get past them. The host user runs everything.

In this section

Approval gateHow risky code and shell commands are scored and approved.Allow-listWho may use the bot.Content screeningChecks on web text, skills and logs.File protectionWhich paths the file tools refuse.SandboxThe Python sandbox: a limited subprocess, not a container.Secret vaultPlaceholders for secrets and how output is cleaned.Threat modelThreats, what stops them, how well, and what is left.

See also:Threat modelApproval gate

Checked against the source at commit f4fd89f on 29 Oct 2026. Everything here is merged on the main branch; no release has been tagged yet. The source is private, so these pages are the public description.