- Data leaves Condense. What a tool returns, such as a pipeline, a configuration or a log, goes to your AI tool, and from there to the model provider that tool uses. Use AI tools that your organization has approved.
- One tool changes your environment.
deploy_connectorcan deploy a connector, so an AI tool that misunderstands a request can deploy something you did not intend.
How Condense limits access
Risks and how to reduce them
Text your AI tool reads can try to steer it
Logs, configurations and documentation are text. A line in a log could contain words written to push an AI tool into doing something you did not ask for. Treat what an AI tool reads from your environment as data, not as instructions, and be careful about what it does next, especially when it can deploy.Recommendations
1
Start with a non-production environment
Connect to a development or test environment first. Choose the environment on the approval page.
2
Use Strict read-only where AI tools should only look
Administrators can set an environment to Strict read-only in AI Governance.
3
Approve only what the AI tool needs
Clear the tools it does not need. If it only reads, clear
deploy_connector.4
Check what it is about to do
Read the tool calls your AI tool asks you to approve, especially a deployment, and read the configuration before you confirm it.
5
Keep secrets out of the chat
Connector configurations can include broker addresses, usernames and passwords. Keep real secrets out of your prompts where you can.
6
Disconnect what you do not use
Remove AI tools you no longer use from the AI Tools page.
For administrators
- Review AI Access Management regularly to see which AI tools members have connected and how much they are used.
- If you suspect a problem, pause first. A pause keeps the connection so you can resume it, and a disconnect makes members connect again.
- Set the operating mode and the allowed tools for each environment in AI Governance before members connect.