The account-wide security settings are on Admin > Account Settings: the automation off switch, how much automation is logged, Guardian, and multi-factor authentication.
Before you start
An admin role. These settings apply to the whole account.
Automation settings
Automation Enabled: the off switch for all automation. When it's off, no automation session can start anywhere in the account.
What each API key may do is set by capability policies, which the link opens. See Create and use capability policies.
Action Log Level: how much automation detail is logged by default. A system can override it, and an API key can raise it but never lower it.
Leave Automation Enabled on normally. If an agent or script is doing something unexpected and you can't find which key, turn it off to stop everything while you look.
Set up Guardian
Guardian is an AI reviewer. For any capability set to Guardian in a policy, it reads the exact action, the agent's reason, what the session has recently done and the open windows, then approves it, denies it, or passes it to a person. It's for work that needs checking on every call, too often for a person to do it.
Each review is sent to the AI provider you choose, using your own key with that provider.
Step 1 - Connect a provider
Turn on Enabled, and use Test to check the connection once the provider is saved.
Provider: for example Google Gemini.
Base URL: the provider's API address. Gemini, OpenAI, Azure, vLLM and Ollama endpoints work. It must be public and use https.
Model.
API key: your key with the provider.
Click Save provider.
NOTE: While Guardian is off, anything set to Guardian asks a person instead, so turning it off is always safe.
Step 2 - Write the mission
The mission is the instructions Guardian follows. It starts with a sensible default.
Always-on rules: rules Guardian always applies, whatever the mission says.
Preset: a starting mission to base yours on.
Mission: edit it to describe your environment: what's normal, what's never acceptable, and what should go to a person.
Below the mission:
Replay against recent decisions runs your edited mission against requests Guardian has already decided, so you can see what would change before you save.
Save mission. Each version is recorded on every decision Guardian makes. Restore default goes back to the original.
Step 3 - Set the limits
Grant lifetime: how long an agent has to make the one call Guardian approved. Guardian approvals are always single-use.
Send screenshots: attaches a screenshot to each review. Off by default, because the screenshot is sent to your provider.
Suspend a key after: suspends an API key after a number of rejections or escalations within a time window. A suspended key is refused until an admin lifts the suspension on the API Keys page.
Click Save limits.
Step 4 - Add context to each key
Each API key has an optional Guardian context, added to the mission when that key's requests are reviewed. Use it to say what the key is for, for example Training class. VMs are disposable. Be relaxed about destructive commands inside the VM, strict about network access. See API keys and their permissions.
Require MFA
Under Security Settings, turn on Require MFA to make everyone in the account sign in with multi-factor authentication. It's the most valuable setting on this page.
Recommended settings
Evaluating in a lab: MFA on; Full access on lab systems; Guardian off.
Production, people only: MFA on; no automation keys, or Read only.
Production with automation: MFA on; Approval for risky; Guardian on, with a mission that fits your environment.
Regulated or customer machines: MFA on; your own policy that restricts more than the risky six; Guardian on, with context on each key.
Check it worked
An agent trying a gated capability produces an approval request instead of acting.
With Automation Enabled off, an agent can't open a session.
Admin > Activity Log records each change, such as
account.security_updatedandcapability_policy.updated.