That is reasonable skepticism. Security actually depends how your model operates and what supervisory or moderation features you layer on top of it regarding the risk posed by prompt injections. That is the first line of defense.
Where I agree with Articul8, is that hosting it internally without having to pass through a LLM or public cloud can reduce the cybersecurity attack surface and the variables listed above can be more tightly controlled. Granted, that doesn't eliminate security risks, just mitigates some of them.
I suspect they're lying about Security with prompt injection being a thing.
That is reasonable skepticism. Security actually depends how your model operates and what supervisory or moderation features you layer on top of it regarding the risk posed by prompt injections. That is the first line of defense.
Where I agree with Articul8, is that hosting it internally without having to pass through a LLM or public cloud can reduce the cybersecurity attack surface and the variables listed above can be more tightly controlled. Granted, that doesn't eliminate security risks, just mitigates some of them.