Fine-tuning takes a general model and teaches it your domain. For security teams that can mean better answers about your own stack, your own policies and your own findings. It also creates a new thing to secure: a set of weights that nobody else has reviewed. Safeguard supports fine-tuning, and it treats the record of how a model was made as part of the product.
What it is
Fine-tuning with Safeguard means adapting a model to your work while the platform keeps track of what went into it. The AI-BOM docs describe the record: every fine-tune you run produces a signed attestation capturing the base model and version, the training data snapshots, the hyperparameters and seed, the evaluation metrics and the environment. Data snapshots are recorded as content hashes, not raw data, so the record does not copy your data.
Lineage follows from that. The docs describe being able to ask, for any served model, which models it was derived from and which datasets touched it.
Why the record matters
- Provenance. A model file is opaque. Knowing where the weights came from, and their hash, tells you whether the file you load is the file you made.
- Load time checks. The loader verifies that the weight hash matches the attested hash and that signatures verify against your trust anchors.
- Data risk. Safeguard flags training data with unknown licenses, copyleft data in proprietary models and regulated data used without controls.
- Audit. When someone asks how a model was built, the answer is a signed document, not a recollection.
Your data stays yours
Safeguard's own models are a separate matter. The model lifecycle page states that its base pretraining excludes customer code at any tier, and that fine-tuning, RLHF and distillation carry the same exclusions, including per tenant prompt history and finding bundles. Customer flagged false positives are excluded unless the customer opted in.
Examples
A team fine-tunes a model on its internal remediation tickets so suggested fixes follow its conventions. The signed attestation lists the base model, the snapshot hashes and the evaluation results. Before the model is served, the loader checks the hash. When an auditor asks where it came from, the team exports the attestation.
Good habits
- Choose the base model with its license in mind.
- Fine-tune on data you are allowed to use, and record it.
- Evaluate on a held out set that includes cases the model should refuse.
- Scan the resulting weights before you load them.
- Keep the attestation with the model, always.
Where it connects
Safeguard also lets you bring your own model, including on-prem and local ones, and run in Zero AI mode with no model at all. A fine-tuned model is one more choice on that list.
Quick check
What does a fine-tune attestation record?
The base model and version, data snapshots as content hashes, hyperparameters and seed, evaluation metrics and the environment.
Why record data as hashes?
So the record can prove what was used without storing a copy of the data.
