Skip to main content

Safety Score

The safety score rates what an agent can actually do: the accounts it uses, what it can change without your approval, and how data could leave. It goes from 0 to 100. Each recommendation comes with a fix you can apply in one click.

The score shows in the agent's header. Click it, or open the Safety tab, to see the details.

Read the score​

GradeScore
A90 and above
B75 to 89
C55 to 74
DBelow 55

90 and above is excellent. 100 is reserved for agents with no access to your accounts or data: an agent that can use your Gmail always keeps a little risk, however well it's set up.

Each item on the tab is rated Weak, Fair, Good or Strong against its own worst case.

Apply recommendations​

Click Apply all recommendations to apply every fix at once. Sprout shows the score you'll reach before you click. Or apply one fix at a time from its row.

Sprout already applies most of these fixes when it builds an agent: tools that change data ask first, high-risk tools start off, and an agent that uses your accounts starts with web tools off. You only see a recommendation if you changed something since, or Sprout couldn't decide for you.

What it checks​

Accounts​

For each connector:

  • Sign-in: an API key loses points when the app also offers OAuth. Fix: Switch to OAuth.
  • Source: a community MCP server loses points when an official one exists. Fix: Use the official one, which keeps the tools that match. MCP servers you added yourself lose a little.

Tools​

For each connector with tools on:

  • Approvals: each tool that changes data or is high risk and runs freely loses points, more for riskier tools. Fix: Ask each time.
  • High-risk tools: each high-risk or critical tool that's on loses points, fewer if it asks first, and fewer again with a rule. Fixes: Add a rule or Turn off.
  • Too many tools: more than 15 tools on a connector loses points. Fix: Review tools and switch off what the agent doesn't need.
  • Access to your account: a small amount of risk that remains whenever an agent can act on your account. You can lower it with rules but not clear it.

A rule on a tool halves what that tool costs.

Web and sandbox​

Willow Agents only. These matter more when the agent holds your accounts, and not at all when it holds no accounts and no memory.

  • Web: reading any web page or searching freely is how data could leave. Fixes: Limit sites, or Turn off.
  • Commands: running commands freely loses points. Fixes: Ask first, or Turn off.

Access bundle​

Instinct and Grok Bot only.

  • Password expiry: a password that lasts more than 30 days, or never expires, loses points.
  • Session length: sessions longer than 30 minutes lose points, if any tool needs a session.
  • Tools that run with the password alone: tools that change data should need a session or ask each time. Fix: Require a session.

Change each setting right in its row. The tab also shows, without scoring them, that only you approve sessions and tool calls, that wrong passwords lock sign-in, and a Rotate shortcut for a new password or link.

Always on​

Some protection doesn't depend on your settings, so it isn't scored. Every Sprout agent has:

  • Prompt-injection shield: messages and tool results that try to rewrite the agent's instructions are blocked.
  • Secrets and personal data masked: API keys, email addresses, phone numbers and card numbers are masked in every tool call.
  • Every tool call is logged: what it ran, when, and with what input. See Logs.

Under Who can use it, the tab also confirms that the agent is private to you and that approvals are yours alone.

On this page