A model file is executable code. Yours arrived unread.
Sense scans every model, adapter and tokenizer before it reaches your registry or your app — payloads, trojaned weights, leaked secrets, unverifiable provenance — and keeps verifying the artifact that is actually running.
You gate every package you install. You gate no model.
A fine-tune downloaded on a Friday afternoon ends up in an app that moves money — with no signature, no bill of materials and no test of what it does when a specific phrase appears in the prompt.
Pickle-based checkpoints execute on deserialisation. One load call on a build machine is remote code execution inside your CI, with your registry credentials in reach.
Base model, training data, fine-tune lineage and licence are asserted in prose by a stranger. Nothing in the pipeline verifies the file is what the card says it is.
Trojaned weights behave perfectly until a trigger phrase arrives. Accuracy metrics look clean, so the only way to find it is to go looking for the trigger.
Five ways a model file betrays the app it ships in.
Pick a finding to see what is inside the artifact, how Sense proves it, and what your team does next.
Twenty checks, five stages, one gate.
Static scanning catches the payload; only behavioural probing catches the backdoor; only signed provenance stops the swap.
Every route into your stack, and what we check on it.
Point us at your registry. We'll tell you what you've already shipped.
A free inventory scan across your model registry and app builds — unsafe formats, unverifiable provenance, licence exposure and backdoor candidates — with findings mapped to the OWASP LLM Top 10 and MITRE ATLAS.