Choose by data sensitivity and the job, not by hype. Zero AI runs without a language model. Bring your own model lets you pick the provider, including local or on premises. Full AI uses the platform's models. The common mistakes are treating Zero AI as a lesser mode and picking a model without reading its terms.
Safeguard offers all three: Full AI, Zero AI and bring your own model, with Claude, GPT, Mistral, Cohere and others pluggable, plus on premises and local options. This post is about the choice itself, with common mistakes in how teams frame it.
Mistake 1: Treating Zero AI as the weak option
Matching a package version against an advisory is a lookup, not a language task. Safeguard's docs describe a Zero model that "answers data questions with no model at all," with no weights and no inference. If your main job is finding known vulnerable components and gating builds, a deterministic mode may do everything you need, and it is predictable. A model adds value for explanation, triage and remediation help.
Mistake 2: Choosing a model without reading the provider's terms
OWASP's LLM03 supply chain entry lists "unclear T&Cs and data privacy policies of the model operators" as a risk, because sensitive data may be used for training and later exposed. Before you connect any provider, read its data retention and training terms, note the date, and record who approved it. Do this again when the terms change.
Mistake 3: Sending everything to the model by default
Models do not need your whole repository to answer most questions. Decide per data class what may leave your boundary: vulnerability metadata, file paths, code snippets, secrets findings. Secrets findings are the obvious one to keep out of prompts. Safeguard's local runner keeps source on the machine and uploads only the SBOM and findings, which is a useful pattern to copy for any workflow.
Mistake 4: One model for every task
Summarizing a finding and writing a patch are different jobs with different risk. A patch that gets applied deserves a stronger model, a human review and tests. A summary can be cheaper and lower stakes. Mixed setups are normal.
Mistake 5: No evaluation before trust
Public benchmarks do not tell you how a model handles your languages and frameworks. OWASP warns that third party models can be fine-tuned to pass published benchmarks. Build a small test set from your own past findings, including cases where the right answer is "this is not exploitable here," and score candidate models on it.
Mistake 6: Forgetting the model is also a dependency
The model, its version and its provider belong in your inventory. A provider can retire a model, change behavior or change terms. OWASP recommends maintained, signed inventories and a patching policy that covers models. Record which model produced which result so you can explain it later.
How should you decide?
- Regulated or restricted data, strict boundary: start with Zero AI or a local model, and add capability only where a gap hurts.
- Need for explanation and remediation help, moderate sensitivity: full AI or a provider you have reviewed.
- Existing provider contract or in-house model: bring your own model, and apply the same review you apply to any vendor.
- Mixed teams: allow different modes per environment, for example Zero AI in the build pipeline and a model in the analyst workspace.
Safeguard's docs also state that your source code, SBOMs and telemetry are never used to train shared models. Read that in context with your own contract, and ask for the data handling terms in writing.
What to do
- List the tasks you want automated and the data each one touches.
- Classify each task as lookup or language work.
- Run lookup tasks in Zero AI first.
- For language tasks, shortlist providers and read their terms, then record the date.
- Build an evaluation set from your own findings.
- Decide what may never be sent to a model, and enforce it technically.
- Add models to your inventory and review them on a schedule.
Quick check
Does Zero AI mean fewer scans?
No. It means the work runs without a language model. Lookup based tasks such as matching components to advisories do not need one.
Can I use my own model with Safeguard?
Yes. Bring your own model is supported, including local and on premises options. Review the provider's terms as you would for any vendor.
Sources
Last reviewed: 18 October 2026.
