Mateo Lafalce on X: "what if they're rigging your API to manipulate you deliberately? Is there any way to know?

with a provider-hosted model: no https://t.co/SBoHEZhzXH" / X<br>Post

Log inSign up

Post

Mateo Lafalce<br>@lafalcemateo

what if they're rigging your API to manipulate you deliberately? Is there any way to know?

with a provider-hosted model: no

span:not(:empty)~span:not(:empty)]:before:content-['·'] [&>span:not(:empty)~span:not(:empty)]:before:px-1 [&>span:not(:empty)~span:not(:empty)]:before:shrink-0">11:42 PM · Jul 28, 202617Views

span:not(:empty)~span:not(:empty)]:before:content-['·'] [&>span:not(:empty)~span:not(:empty)]:before:px-1 [&>span:not(:empty)~span:not(:empty)]:before:shrink-0 min-w-0 overflow-hidden">Mateo Lafalce<br>@lafalcemateo

21h

Through an API you're blind to the three layers that matter:

- which model is actually answering<br>- what the effective system prompt is<br>- whether the weights were modified

You're entirely at the provider's mercy

svg]:size-5 text-body hover:bg-mix-current hover:bg-mix-amount-10 active:bg-mix-current active:bg-mix-amount-15 focus-visible:bg-mix-current focus-visible:bg-mix-amount-10 outline-current -m-2 shrink-0 cursor-pointer border-transparent p-0 text-body size-9 [&>svg]:size-[1.25em] [&>[data-engagement-icon]]:size-[1.25em] group-hover:bg-mix-current group-hover:bg-mix-amount-10" aria-label="View count" type="button" data-state="closed" href="/lafalcemateo/status/2082250307950592056/quotes">8

span:not(:empty)~span:not(:empty)]:before:content-['·'] [&>span:not(:empty)~span:not(:empty)]:before:px-1 [&>span:not(:empty)~span:not(:empty)]:before:shrink-0 min-w-0 overflow-hidden">Mateo Lafalce<br>@lafalcemateo

21h

Running inference on your own VPS gets two of them back: you can validate the model version and the system prompt.

But the weights are the Achilles' heel. It's a file written once that nobody ever looks at again, and whoever owns the box has access.

svg]:size-5 text-body hover:bg-mix-current hover:bg-mix-amount-10 active:bg-mix-current active:bg-mix-amount-15 focus-visible:bg-mix-current focus-visible:bg-mix-amount-10 outline-current -m-2 shrink-0 cursor-pointer border-transparent p-0 text-body size-9 [&>svg]:size-[1.25em] [&>[data-engagement-icon]]:size-[1.25em] group-hover:bg-mix-current group-hover:bg-mix-amount-10" aria-label="View count" type="button" data-state="closed" href="/lafalcemateo/status/2082250310475526634/quotes">8

span:not(:empty)~span:not(:empty)]:before:content-['·'] [&>span:not(:empty)~span:not(:empty)]:before:px-1 [&>span:not(:empty)~span:not(:empty)]:before:shrink-0 min-w-0 overflow-hidden">Mateo Lafalce<br>@lafalcemateo

21h

In the ideal privacy world you run the model on your own hardware: model, prompt and weights all legitimate and hash-verifiable.

That's the practical ceiling.

svg]:size-5 text-body hover:bg-mix-current hover:bg-mix-amount-10 active:bg-mix-current active:bg-mix-amount-15 focus-visible:bg-mix-current focus-visible:bg-mix-amount-10 outline-current -m-2 shrink-0 cursor-pointer border-transparent p-0 text-body size-9 [&>svg]:size-[1.25em] [&>[data-engagement-icon]]:size-[1.25em] group-hover:bg-mix-current group-hover:bg-mix-amount-10" aria-label="View count" type="button" data-state="closed" href="/lafalcemateo/status/2082250313453494536/quotes">6

span:not(:empty)~span:not(:empty)]:before:content-['·'] [&>span:not(:empty)~span:not(:empty)]:before:px-1 [&>span:not(:empty)~span:not(:empty)]:before:shrink-0 min-w-0 overflow-hidden">Mateo Lafalce<br>@lafalcemateo

21h

Even then you're not safe: the software and hardware you run can carry backdoors. Firmware, drivers, microcode, silicon.

Trust doesn't disappear, it moves. The question is never "do I trust?" but "whom, and at which layer?"

New to X?<br>Sign up now to get your own personalized timeline!<br>Sign up with GoogleSign up with AppleCreate account<br>By signing up, you agree to the Terms of Service and Privacy Policy, including Cookie Use.

Relevant people

Mateo Lafalce@lafalcemateoFollow<br>Neurons that fire together, wire together 𖣂

Trending now

Don't miss what's happening<br>People on X are the first to know.

Log inSign up