Before rolling out an AI coding assistant, ask the vendor where code and prompts go, how long they are kept, whether they train models, what admins can control, how agent features are limited, and how incidents are reported. Get the answers in writing, then test the claims on a pilot team.
Data handling
Ask what leaves the developer's machine: open files, repository context, terminal output, telemetry. Ask where it is processed and stored, and for how long. Ask whether your data is used to train or improve models, and how that is switched off for your account rather than only described in a policy page. Ask for the subprocessor list and data residency options if you need them.
Admin and policy controls
Look for the ability to enforce settings centrally: which models are allowed, whether agent mode is on, which tools and servers can be connected, and whether individual developers can override. Ask about audit logs. Can you see who used the assistant, on which repositories, and what actions an agent took? If controls exist only as per-user toggles, you do not have enforcement.
Agent features and extensions
Assistants increasingly run commands, edit files and call external tools. The OWASP Top 10 for LLM Applications lists prompt injection, sensitive information disclosure, supply chain and excessive agency among its entries, and all four apply here. Ask how the product limits what an agent can do by default, how approvals work, and how extensions or plugins are vetted. Ask how it treats content from repositories or web pages, since that content can carry instructions.
Security of the vendor
Ask for the independent assurance reports the vendor can share, a vulnerability disclosure process, and a history of how past issues were handled and communicated. A calm, specific answer matters more than a clean record. Finally, ask about exit: can you delete your data, and can you confirm it.
Verify, do not trust
Put a small pilot group on the product with a written checklist. Check outbound connections with your own tooling. Try a repository file that contains an instruction aimed at the assistant and see what happens. Compare behavior with the vendor's description.
Some questions are about people rather than software. Who at the vendor answers a security escalation, and how fast? Is there a named contact or only a ticket queue? Ask also how the vendor communicates changes to data handling terms, since policies on retention and training use can shift. Subscribe to those notices, and assign someone on your side to read them. A good control today can quietly weaken after a terms update that nobody noticed.
Where Safeguard fits
Safeguard offers policies and guardrails, an MCP Guardrail and agent identity that can sit around assistants and agents, and SBOM and dependency scanning for the code they produce.
Quick check
Should I block AI assistants until the review is done?
That is a business choice. Many teams run a small, monitored pilot instead, which gives better evidence than a paper review alone.
Sources
Last reviewed: 19 October 2026.
