Safety controls

The scope is the hard boundary: agents can never reach an out-of-scope host. Within the scope, these controls decide how much the agents may do without asking you, and how you stop them.

Permission modes

Every project has a Permission mode. It decides which agent actions need your approval.

Mode	Behaviour	Use it when
Bypass approvals	In-scope actions run without asking.	You trust the scope and want an uninterrupted run, typically on a staging system.
Auto-classify actions (default)	Agents run normal actions on their own. An action that the model's safety classifier flags as risky is held until you decide.	The usual choice: you are only asked about risky actions.
Approve flagged actions	The strictest mode: flagged actions always wait for your decision.	Sensitive targets where you want to review risky actions yourself. Expect many approval requests.

In every mode, an action that targets an out-of-scope host is never executed silently; it becomes an approval request.

Depending on the model, Auto-classify can still produce many approval requests (models differ in how much they flag). If that happens on a staging system, switching to Bypass approvals is usually the right answer.

Bypass approvals is your decision

With Bypass approvals, nobody reviews the agents' actions before they run. Every in-scope action is executed, including actions that change or delete data, lock accounts or put load on the system. Choosing this mode means you accept the impact of all those actions on the target, just as if you had approved each one yourself.

Only choose it if the system owner has agreed to this kind of testing and the target can tolerate it, typically a staging system that can be reset. The scope boundary still applies: out-of-scope hosts remain unreachable.

Changing the mode during a run

You can change the mode at any time on the Overview tab (Apply mode).

• Loosening the mode (for example to Bypass approvals) applies immediately, also to agents that are already running.
• Tightening the mode applies to agents that start afterwards. Agents that are already running keep the looser mode until they finish. Pause and resume the project if the stricter mode must apply to everyone at once.

Approvals

When an agent needs your decision, the workspace shows an approval request with the agent, the action and its reason.

• Allow once lets this single action run.
• Allow & add host (for a host that is out of scope) lets the action run and adds the host to the project's scope permanently. Only use it if your authorization covers that host.
• Deny blocks the action. The agent continues with something else.

An approval request that nobody answers is denied automatically after about 10 minutes. Only that single action is denied; the run continues.

You are responsible for what you allow

An approval is a decision you make on behalf of the test. When you allow an action, you take responsibility for its effect on the target, including data changes, locked accounts, triggered emails and load. When you add a host, you confirm that your authorization covers it.

Read the action and its reason before you decide. If you cannot judge the impact, choose Deny: the agent continues with other tests, and you lose at most that single action.

The Approvals button in the workspace header shows all past decisions, including who decided and when.

Pause and Resume

Pause stops all agents of the project. Resume continues the run where it stopped; agents pick up their work where they left off. Use Pause whenever you need the agents to stop for a while, for example when the system owner reports a problem.

Stop hunting

Stop hunting ends the testing for good. Use it when the time box is over.

• Hunting agents that are running or waiting are aborted.
• Every test that has not run yet is marked as skipped.
• Findings that already exist are still verified, reported and illustrated.
• The project then finishes as completed with gaps.

Stop hunting cannot be undone

Hunting cannot be resumed after Stop hunting. If you only want to stop for now and continue later, use Pause instead.

If something goes wrong on the target

If the target shows problems during a run (errors, high load, unexpected data changes):

1. Press Pause. All agents stop.
2. Check the Agents tab to see which agents were running and what they did.
3. Clarify the problem with the system owner. If needed, narrow the scope with Edit scope or tighten the permission mode.
4. Press Resume to continue, or Stop hunting to end the test with the findings found so far.

Checklist

• I chose a permission mode that fits the target (staging or production).
• If I use Bypass approvals, I accept the impact of every in-scope action the agents take.
• I only allow actions whose impact I can judge and accept.
• I will be available to answer approval requests during the run, or I chose Bypass approvals.
• I know the difference between Pause (resumable) and Stop hunting (final).