Skip to main content

Threat model

Knox Agent runs tools on your machine. The attacker does not control you or the model provider. The attacker controls text the agent reads and tries to turn the model into a confused deputy: read a secret, run a command, write outside the project, or leak data over the network.

This page is a user abridgement of the 2.2 threat model. It does not list source file paths.

Trust​

TrustedUntrusted
Your own promptFile contents (cloned repo, dependencies)
User-level config (~/.knoxcoder)Web pages and search results
What you explicitly ApproveTool output (logs, tests, git messages)
AGENTS.md, skills, custom agents, hooks, team bundles, subagent output

Every untrusted input can become model-visible text. Knox does not try to detect injection in the prose. It limits what an injected instruction can do, in the tool layer.

What Auto allows (by design)​

Auto auto-approves enabled tools. A hostile repo can make the agent change the project’s own code. Review the diff (Review edits, checkpoints) before you run or commit.

Fresh clones start in Ask. Stay in Ask or Edits on trees you do not trust. Treat Auto on an untrusted repo as a disposable environment.

Controls (always on)​

  • Path/command policy; deny always wins
  • Hard path denies (~/.ssh, ~/.gnupg) on lexical and resolved paths
  • Secret files always ask, even in Auto
  • Workspace boundary (symlinks resolved)
  • Destructive / download-and-execute command guard — Command guard
  • Subagent merge cannot escape the trees
  • builtin_fetch_url blocks private ranges per redirect hop
  • Secret redaction on tool output and Memory Brain
  • Knox does not activate in an untrusted VS Code workspace, so repo hooks never run on folder open
  • Team bundle import skips hooks unless you allow them

AGENTS.md / skills / custom agents are prompt text. A custom agent’s tools: list can only narrow, not widen.

Open holes (be honest)​

  • Encoded scripts, make recipes, deeper indirection past the command guard
  • DNS rebinding on fetch
  • Leakage to a public URL via query string
  • Instructions in tool output still reach the model
  • Recipe replay is a code-execution surface (experimental, off) — mitigated, not reviewed as a dedicated audit

Next​

Permissions and command guard · Permission modes · Known limitations