Skip to content
Docs menu

Security

How venom decides what the agent may do, keeps secrets out of the model, contains commands in an OS sandbox and leaves evidence you can verify offline.

View .md

Venom runs a model that can execute commands and edit files on your machine. The controls in this section are layered so that no single one has to be perfect. Each one works offline with a local model.

Layer What it does Page
Permission modes and rules Every tool call is judged before it runs: hard deny, deny rules, ask rules, allow rules, then the mode's default. Permissions
OS sandbox Bash runs under Seatbelt (macOS) or bubblewrap (Linux): writes stay in the workspace and temp dirs, network goes through an allowlist proxy. Sandbox
Project trust A cloned repository can't turn on its own hooks, MCP servers or allow rules until you trust it. Project trust
Organisation policy An administrator's file, or a signed policy from Venom Cloud, caps modes, rules, MCP servers and the provider for everyone on the machine. Organisation policy and risk
Secrets vault The agent writes {{secret:NAME}}; the value is put in at run time and scrubbed from every result. Secrets vault
Guardrails Content, routing, model, tool, spend and answer rules at the org, key group and member layer, each with a stable code. Includes the prompt-injection shield and the supply-chain check. Guardrails
Session guard and runaway guards A session that loops, stalls or overspends is paused, not killed. Session guard
Audit trail A hash-chained permission audit log, a signed log of every command and write, and exports that verify offline. Audit, signed actions and compliance
Egress map Which hosts each session contacted, with new and off-allowlist hosts flagged. Egress map
Red team A versioned attack corpus run against your guardrails and a gear. Red team

Where to start

What venom does not claim

A policy binds the venom process a user runs, not the user: anyone with a shell can run another program. The heuristics (command parsing, secret-leak checks, injection detection) are defence in depth, not proofs. The OS sandbox and the egress allowlist are the boundaries that still hold when a model is talked into trying something.

In this section

  • Permission modes and rules: Choose a permission mode, write allow, ask and deny rules, and understand the order in which venom judges every tool call.
  • OS sandbox and network allowlist: What sandboxed bash can read, write and reach on macOS and Linux, and how to change its network allowlist.
  • Project trust: Decide whether a project's own hooks, MCP servers, allow rules and loosening settings may run, from the TUI dialog or headless with venom trust.
  • Organisation policy and risk tiers: Enforce rules for every venom on a machine with /etc/venom/policy.json or a signed policy from Venom Cloud, and score the risk of a change with venom risk.
  • Secrets vault: Store tokens once with venom secrets set, let the agent use them as {{secret:NAME}} without ever seeing the value, and bind each one to the origins it may reach.
  • Guardrails: Set content, routing, model, tool, spend and answer rules at the org, key group and member layer, read a guardrail code, and use the prompt-injection shield and supply-chain check.
  • Session guard and runaway guards: Pause sessions that stall, loop or overspend, set time and dollar caps with session_guard, and set an output budget with guards.
  • Audit log, signed actions and compliance: Verify the hash-chained permission audit log, the signed log of every command and file write, offline audit exports and Venom Cloud compliance packs.
  • Network egress map: See which hosts each session contacted through bash, the browser, web tools and MCP, with hosts outside the allowlist or new for the repository flagged.
  • Red team: Run the versioned attack corpus against your guardrails and a gear with venom redteam run, within a spend cap, and read which guardrail caught each case.