Choose an AI mode by asking three questions: what data will the model see, who operates the model, and what is it allowed to do. Zero AI suits strict data rules. Bring your own model suits teams that already govern a model. Full AI suits teams that want the fastest help.
What are the three modes?
Safeguard offers three. Full AI runs Safeguard's models in the loop. Zero AI runs the platform without any language model, so the scanners work deterministically. Bring your own model lets you plug in a model you choose, including Claude, GPT, Mistral and Cohere, or one hosted on premises or locally.
Other security vendors use similar labels, and the checklist below works for any of them.
Which questions should drive the decision?
What data reaches the model? OWASP's list of risks for language model applications includes sensitive information disclosure. Source code, findings and customer records can all be sensitive. Map what a prompt would contain before you choose a mode.
Who runs the model, and under what terms? A model run by your own team sits inside your controls. A hosted model sits inside the provider's terms. Read the retention and training terms for the exact service tier you would use, not the marketing summary.
What may the model do? Reading and summarizing is lower risk than writing code, opening pull requests or calling tools. OWASP advises least privilege and human approval for high risk actions. The NCSC adds that a model processing outside content is confusable, so tool access should be limited to what an attacker could safely be given.
Can you explain a result later? NIST's AI Risk Management Framework asks for monitoring in production and documented limits. If an assessor asks why a finding was ranked first, you need a record, not a vibe.
What does failure look like? If the model is unavailable or wrong, does work stop, or does the deterministic path carry on?
How do the modes compare in practice?
Zero AI gives the most predictable behavior and the smallest review surface. You give up natural language help, such as explanations and drafted fixes, so people do more by hand.
Bring your own model gives you control over provider, region and terms, and lets you reuse a model your organization has already approved. You take on the work of governing that model, including access, logging and updates.
Full AI gives the least setup and the most help per minute. You depend on the vendor's model hosting and need to review its data handling terms against your policy.
Regulated teams often mix modes by environment, for example Zero AI for the most sensitive repositories and a governed model elsewhere. Ask whether the setting can be changed per workspace rather than only per company.
What to do
- Classify the data each AI feature would send, and who can see the output.
- Write your rule for which classes may reach which kind of model.
- List what the model may do on its own and what needs human approval.
- Read data retention and training terms for the exact tiers in play.
- Pilot the mode on one team, and compare time saved against review effort added.
- Record the decision, the owner and a review date, then revisit when models or rules change.
Quick check
Is Zero AI less capable at finding vulnerabilities?
Scanning itself can run without a model. What you lose is language assistance, such as explanations and drafted fixes. Test your own repositories to see what matters to you.
Does bringing our own model mean our data stays private?
Not automatically. It depends on where that model runs and on the provider's terms. A model hosted inside your environment gives you the most control.
Sources
- Safeguard docs, fetched 18 October 2026
- OWASP Top 10 for LLM Applications, fetched 18 October 2026
- OWASP LLM01: Prompt Injection, fetched 18 October 2026
- NCSC: Prompt injection is not SQL injection, fetched 18 October 2026
- NIST AI RMF Core, Measure and Manage functions, fetched 18 October 2026
Last reviewed: 18 October 2026.
