Mistral Security Boundary
Overview
Apply controls at the application boundary because model instructions do not enforce identity, tenancy, authorization, or side-effect safety. Treat user content, retrieval, files, and output as untrusted.
Prerequisites
- A data-flow diagram covering browser, app, Mistral, stores, and tools.
- Credential, retention, tenant, moderation, and incident policies.
- An inventory of stateful APIs, files, retrieval stores, and actions.
Current Contract
Bearer credentials authorize provider calls but not application users. Classifiers can support policy but do not replace controls. ZDR eligibility differs between stateless and stateful products.
Authentication
Keep keys server-side and environment-scoped; rotate on exposure. Enforce user and tenant authorization before retrieval, provider calls, files, and tools.
Instructions
- Map trust boundaries and classify prompts, outputs, embeddings, files, logs, and IDs.
- Remove credentials and privileged instructions from browser/untrusted contexts.
- Separate trusted policy from untrusted content and delimit retrieved text as data.
- Validate output structure and safety before render, storage, or tool use.
- Authorize tools via closed registry, schemas, least privilege, deadlines, and idempotency.
- Test injection, cross-tenant access, unsafe output, exfiltration, duplicates, and rotation.
Tool Discipline
Use Read, Glob, and Grep to inspect code, locks, configuration, tests, and evidence. Use Write and Edit only for approved repository changes. Invocation alone does not authorize network calls, paid usage, uploads, stateful resources, admin mutations, deployments, or deletion.
Approval Boundaries
Require approval for sensitive data, uploads, stateful APIs, tools, policy changes, retention, or live attack testing. Model output never grants authority.
Error Handling
- Prompt wording cannot enforce access control.
- Moderation success does not prove correctness or tool safety.
- Assuming ZDR for Files, Batch, Agents, Conversations, or other stateful products is unsafe.
Output
Return assets, threats, controls, residual risks, data/API boundary, tests, owners, and rollback. Exclude live secrets and customer data.
Examples
- Reject retrieved text asking for a system secret.
- Require fresh app authorization before a proposed refund tool.
Validation
Run synthetic abuse cases across input, retrieval, output, rendering, and tools; confirm tenant filters and rotation independently.
Resources
- Current first-party evidence map — recheck dated sources before relying on mutable endpoints, models, limits, prices, preview status, or retention.
- Record live account observations as environment-specific evidence, not universal Mistral guarantees.
