Depends on how you're running them I suppose. If you are running them on local hardware, then the harness is what you are concerned about. A compromised local model might be able to exfiltrate data via tool calls.
If you are running it in a cloud, the above is true, but also you have to trust the cloud provider isn't relaying your chats to a third party, for training or perusing.
If I am totally blunt, if privacy is a concern then corporate US has proven time and time again to be a terrible steward of your private data. You should assume it's being ingested or sold, or turned into metadata that's sold in a data laundering pipeline.
If it's corporate secrets, on-prem and enterprise offerings are your options. Enterprise at least puts the providers on the hook legally, when they inevitably fuck up keeping your data private in some way.
Indeed - if you need to be absolutely confident about security and privacy, run a model locally and audit the inference software and potential tooling.
If you are running it in a cloud, the above is true, but also you have to trust the cloud provider isn't relaying your chats to a third party, for training or perusing.
If I am totally blunt, if privacy is a concern then corporate US has proven time and time again to be a terrible steward of your private data. You should assume it's being ingested or sold, or turned into metadata that's sold in a data laundering pipeline.
If it's corporate secrets, on-prem and enterprise offerings are your options. Enterprise at least puts the providers on the hook legally, when they inevitably fuck up keeping your data private in some way.