It depends on the feature, and you should ask rather than assume. Some AI features read only metadata such as package names and versions. Others read source files, diffs, or tickets. Ask the vendor to document, per feature, exactly what data reaches a model, where it is processed, and how long it is kept.
Three levels of exposure
| Level | Example inputs | Typical risk |
|---|---|---|
| Metadata only | Package names, versions, CVE ids | Low, reveals your stack |
| Findings and context | Alert text, file paths, snippets | Medium, may include sensitive names |
| Full source or diffs | Whole files, commits, comments | Higher, may include secrets or proprietary logic |
Ask which level applies to each AI feature separately. A vendor may use one level for triage and another for fix generation.
The data path questions
- Is the data sent to a model processed by the vendor, or by a third party?
- Is it stored, and if so, for how long and where?
- Is it used to train or tune any model? Ask about prompts and outputs separately.
- Can you exclude repositories, paths, or file types?
- Is there an option to keep processing inside your environment?
- What logs exist, and can you read them?
Secrets in code
Source files sometimes contain credentials. If a feature sends file contents to a model, a secret in a file may travel with it. Ask what the vendor does to reduce that exposure and, separately, clean up your repositories regardless of which vendor you choose.
Read the written terms
Marketing pages summarize. Terms and data processing documents govern. Read both, and have your privacy and legal teams review the data processing terms, since this post is not legal advice. When the page and the document disagree, trust the document and ask for a correction to the page.
Guidance worth citing internally
CISA, together with the NSA, the FBI, and international partners, published guidance in May 2025 on securing data used to train and operate AI systems, covering data risks across the lifecycle. Pointing to it can help justify a data handling review to colleagues who see it as a delay.
Try it yourself
Click through it at your own pace, or open it full screen.
Notice how much of an inventory is names and versions rather than source. Ask whether a feature needs more than that.
Where Safeguard fits
Safeguard provides SBOM and dependency scanning, and a free scan at gold.safeguard.sh/scan. Ask us the same data path questions for any feature you consider.
Quick check
Can I opt out of AI features?
Ask. A mature vendor lets you disable features per team or project.
Is metadata only really safe?
Safer, not harmless. A list of packages reveals your stack and its weaknesses.
Sources
- CISA, AI Data Security: Best Practices for Securing Data Used to Train and Operate AI Systems, published 22 May 2025: https://www.cisa.gov/resources-tools/resources/ai-data-security-best-practices-securing-data-used-train-operate-ai-systems
Last reviewed: 19 October 2026.
