Security and privacy
Security and privacy
Bots read your apps and act in them, so set them up the way you would a new colleague's access: only what the job needs, with a check on anything that can't be undone.
Who can see what
- Your chats are yours. On a team, owners and admins can't read them, and a teammate using the same shared bot has a separate chat with it.
- A shared bot's settings can be seen and changed only by the person who made it and the team's owners and admins.
- Pages are private until you share them. See Pages.
- Knowledge sources answer everyone in their space through the account of the person who added them. See What's shared in a team space.
- Opisense staff on Opisense Cloud can see an organization's members, spaces, bots and skills to help with support, but not its chats, memory, files or saved credentials. Changes they make are recorded in the organization's audit log.
By default, Opisense encrypts chats, memory, files and credentials before storing them, so a copied database or backup doesn't reveal them.
Check actions before they happen
Approvals are the safety net, whatever a bot's instructions say:
- Ask me for a bot that sends, posts, pays or deletes, until you trust it. Auto review lets through actions that fit what you asked and asks about ones that go further.
- A Block rule in Settings → Approvals for anything that must never happen.
- Disable a tool in a bot's Advanced settings to take a tool away from that bot.
Cards show what the bot is about to do, such as the email as it will be sent.
Outside content
Email, web pages, webhooks and messages from other people can contain text written to steer a bot. Opisense limits that:
- In Ask me and Auto review, a routine started by a webhook, a Git event or a Slack message needs your yes for every action that changes something; the approval model can't pass it.
- Once a run has read outside content, even a read from a new site or app goes through approval.
- Runs you didn't start yourself, such as a message from a chat app or another bot, can't change a bot's instructions or model without you.
- Never ask turns most of these checks off, for every bot in the account or space it's set for. Don't use it where bots read outside content.
The risky combination is a bot that reads private data, reads content from strangers, and can send things out. Remove one of the three: keep approvals on for sending, or give inbound mail and web research to a bot whose sending tools are disabled, and let it report to you.
Credentials
- Save keys and passwords through the protected card a bot shows, or in Settings → Vault. Never paste them into the chat.
- The model never sees saved values, and a saved login is typed only on its own site.
- A Command variable is readable by the bot's own shell commands. Save one only for a value you would let that bot use unattended.
- Use read-only or narrowly scoped keys when the service offers them.
- Give a bot that signs in to sites a Private computer, so other bots don't share its sign-ins.
See Reusable credentials.
Your own computer
The desktop app can let bots run commands on the computer it's installed on. A bot with that access can read your files and run programs as you. Turn it on only for a computer you would let your bots use unattended.
Sharing outside your team
- Public page links let anyone with the link open a page. Owners and admins can turn them off for the whole team in Settings → General.
- In a Slack room, anyone in the room can wake an approved bot, and it acts with the account and apps of the person who installed it. Use Mentions only and keep approvals on. See Chat apps.
- Sharing a bot to the Marketplace shares only what's on the form, never chats, memory, credentials or connected accounts.
Your data
- Settings → General → Export data downloads your chats, files and bot workspaces.
- Chats and memory stay until you delete them: Clear conversation empties a chat, and a note can be deleted from Memory. Delete account removes everything you own.
- For a DPA, security documents and the list of subprocessors, see the Trust Center.
- To report a security issue, email support@opisense.com.
- For full control over where data and models run, host Opisense yourself (contact us for setup).