ZeroTwo has no separate security-research models or access tier, so authorized security work uses the same agent and the same controls as any other task. The controls you need grow with what the agent can reach: a lab repository needs little, while anything touching production systems, sensitive data or outside tools calls for the stricter measures below.
You own the risk assessment. Approvals and Auto-review reduce risk but can make mistakes, and they do not replace your own security, monitoring and oversight.
Assume the model, its tools and connected systems could be compromised, and set things up so they still cannot reach unauthorized targets, leak credentials or persist.
Isolate the environment
Prefer a cloud environment for risky work. Cloud tasks run in a managed sandbox, so the virtual machine rather than an approval prompt is the boundary. In Settings > Code Cloud, each environment has an Allow ZeroTwo to access the internet switch; with it on, Allow domains offers Package managers, Custom domains only and All (unrestricted). Choose Custom domains only and list just your lab hosts, or switch internet access off.
Local runs are weaker. On macOS, local commands run under an OS sandbox. Windows and Linux have none, so approvals are the only safeguard there; use a dedicated machine or virtual machine for anything untrusted.
Define and enforce approved boundaries
Write down in-scope and excluded systems, allowed and prohibited actions, the time window, data handling and stop conditions, and put them in a project instruction file such as AGENTS.md. That file is context for the agent, not enforcement, so back it with controls ZeroTwo applies itself:
- Choose a permission preset on desktop: Ask for approval lets the agent edit its workspace and ask for more, Approve for me adds automatic review, and Full access removes approvals.
- Use Plan mode to have the agent research and propose steps first; the turn runs read-only.
- Set each connector or MCP tool to Always allow, Needs approval or Blocked, and remove connectors the task does not need.
- On desktop, add a trusted
PreToolUsehook under Settings > Coding > Hooks to deny tool calls that fall outside the approved scope. - Leave Let ZeroTwo control the built-in browser and Computer use off unless the task needs them.
Protect credentials and sensitive data
Keep long-lived keys out of prompts, repositories, instruction files and memory. Memory redacts obvious secrets, but do not depend on it. Prefer short-lived, target-scoped credentials, give the task only the data it needs, and treat generated files as untrusted.
In Settings > Code Cloud, the personal vault can hold a secret with Allowed domains, so only requests to those HTTPS domains can use it.
Avoid Full access. The desktop warns that it can edit any file on your computer and run commands with network access without your approval. Local scheduled tasks run with Full access and cannot ask, so do not schedule risky work locally.
Review sensitive actions
Require a human decision before irreversible changes and before anything touching production, external systems, persistence, privilege escalation or sensitive data. Keep Ask for approval for these. Approval cards offer Yes, Yes, and don't ask again this task, an always option, and No with feedback, and a pending approval can be answered from the iOS or Android app. Be sparing with the always option, since it saves a permission rule.
Approve for me, called Auto-review in Settings, sends requests for extra access to an automatic reviewer instead of you. It is desktop only, it can make mistakes, and it is not a substitute for a human on high-risk steps. Ignore instructions embedded in web pages or tool output.
Monitor independently
ZeroTwo has no audit log or organization-level activity export, so monitor from outside it: log network and credential use where the model cannot reach, alert on off-target activity, and keep your own emergency shutdown for the environment. Stop the task if a critical control fails.
Clean up and validate the results
Revoke temporary credentials, remove environment variables and secrets you added, stop processes and delete risky environments. Validate findings, follow coordinated disclosure, and keep people accountable for fixes.
Before you start
Confirm scope, isolation, minimal permissions, a restricted network, protected credentials, action review, independent monitoring, an emergency stop and cleanup. No single control should stand alone.