Security

Last updated 29 October 2026

An agent that browses the web for you needs guardrails. Here are ours, in plain language, including the limits.

The short version

  • Each agent gets its own isolated cloud browser and sandbox.
  • Secrets are encrypted with AES-GCM and never shown to the AI.
  • Connections are default-deny: an agent can reach only the hosts you allow.
  • Emailing people outside your workspace or writing to other services asks you first. On Free, always.
  • No browser agent is immune to prompt injection. We limit what a tricked agent could do.

Isolation

  • Each agent works in its own cloud browser and its own code sandbox on Cloudflare’s infrastructure, never in your own browser and never on your computer.
  • An agent’s browser profile, meaning its cookies and sessions, belongs to that agent alone. Another agent can’t use its logins.
  • Workspaces are walled off from each other. Every request is checked against your membership, and roles (owner, admin, member) decide who can change agents, connections and billing.
  • Published pages are served from a separate domain, apps.gentlebot.ai, so a page can’t reach your signed-in session. Files your agents make are displayed in a sandbox with no access to your account.

Secrets and logins

  • Connection tokens and API keys are encrypted at rest with AES-256-GCM.
  • They’re decrypted only on our servers, at the moment a request is made, and added to that request directly. The AI model never sees them, and the app never shows them again after you save them.
  • When a site needs a login, you take over the agent’s live browser and sign in yourself. The session is kept in that agent’s profile; we don’t save the password you type.
  • Signing in to Gentlebot requires a verified email address, and passwords are stored only as salted hashes.

Network limits

  • Connections are default-deny. Each one lists the hosts it may reach, and requests anywhere else are refused.
  • Requests to private and internal network addresses are blocked, to prevent server-side request forgery (SSRF). The tools that fetch web pages use the same guard.
  • Tools from connected apps (MCP servers) stay off until a workspace admin allows them. We record each tool’s definition when it’s approved; if a server changes a tool later, that tool is set aside until an admin reviews the change.

What asks first

This is the full list of actions that wait for your approval. Trusting an agent is a per-agent setting on Pro and Max; on Free it isn’t available.

ActionFreePro and Max
Emailing anyone outside your workspaceAlways asksAsks, unless you trust the agent
Sending data to another service through a connection (anything beyond reading)Always asksAsks, unless you trust the agent
Write actions in connected apps (tools that aren’t read-only)Always asksAsks, unless you trust the agent
Creating or changing its own schedule or triggersAsksAsks, unless you trust the agent
Downloading a file from the web into its sandboxAsksAsks, unless you trust the agent
Giving an agent access to one of your connectionsAsksAsks, unless you trust the agent
Changing its own AI model, raising its budget, turning on a new tool, or asking for a new login or keyAlways asksAlways asks

Never asks, because it’s the job: opening and reading pages, searching the web, running code in its own sandbox, saving files, building and updating its pages, remembering things, and emailing you or your teammates. Making a page public is always your decision.

Inside the browser Clicks and typing aren’t approved one by one. The agent is instructed to stop before it posts, buys or changes an account, and to hand you a draft or the controls. That’s an instruction to the model rather than a hard switch, which is why the isolation and network limits above exist, and why we suggest logging an agent in only to the sites its job needs.
When nobody is there to ask Runs started by a schedule, a webhook or an incoming email stop at an approval and wait for you. A few runs can’t wait: those started from Slack, or by a button on one of your pages. On Free, outward actions in those runs are refused and left as drafts. On Pro and Max they go ahead, so set those triggers up only for agents you trust.

Prompt injection, plainly

A web page, email or file can hide instructions meant to hijack an AI agent. This is called prompt injection, and no browser agent is immune: not ours, not anyone’s. So instead of promising it can’t happen, we limit what a tricked agent could do.

  • Your secrets aren’t there to steal. Keys and tokens never enter the model’s context.
  • It can’t send your data just anywhere. Connections reach only the hosts you allow, and private networks are blocked.
  • It can’t act for you without asking. Emails to people outside your workspace and writes to other services need your approval, always on Free.
  • It can’t wander into your other accounts. Each agent’s browser holds only the logins you gave that agent.
  • You can see what it did. Every run has a receipt, and you can watch the browser or take it over at any time.

Smaller, faster models are generally easier to mislead. That’s one reason Free, which runs on Claude Haiku, is always draft-only.

What you can do

  • Give each agent only the logins and connections its job needs.
  • Keep approvals on for agents that read content you don’t control, such as inboxes, marketplaces and forums.
  • Read drafts before you approve them.
  • If you work for clients, use a separate workspace for each one.

Protecting your data

  • Data is encrypted in transit (TLS) and at rest in Cloudflare’s storage.
  • Every API request is checked for a valid session and workspace membership, and every agent, page and file is checked against the workspace that owns it.
  • Payments are handled by Polar, so we never see your card number.
  • Your content goes to AI providers only to do the work, under commercial terms that don’t allow training on it. Our privacy policy lists who processes what.

Deleting data

  • Deleting an agent removes its chats, memory, files, browser session and pages.
  • Deleting your account (Settings → Account → Delete account) removes every workspace you own and everything in it, and cancels any paid plan. Here’s exactly what’s removed and what’s kept.
  • Deleted data can remain in encrypted backups for up to 30 days.

Reporting a vulnerability

Found a security issue? Email security@gentlebot.ai with the steps to reproduce it, what an attacker could do with it, and any proof of concept. Please:

  • give us reasonable time to fix it before telling anyone else;
  • use only your own accounts and data, and don’t access, change or delete anyone else’s;
  • don’t run denial-of-service tests, send spam, or try to trick our team or users.

If you act in good faith and follow these rules, we won’t take legal action against you, and we’ll credit you publicly if you’d like. We aim to reply within 3 business days. We don’t run a paid bug bounty yet.