Microsoft Foundry SDK for Python 2.5

Apply the Python 3 baseline first. These rules cover Microsoft Foundry SDK for Python 2.5 project endpoints, Azure identity, connections, model clients, project assets, evaluations, tracing, networking, and operations.

Project client and identity

• Construct AIProjectClient with a fixed trusted project endpoint and an explicitly selected production credential.
• Prefer ManagedIdentityCredential or WorkloadIdentityCredential in production; if DefaultAzureCredential is required, exclude developer and interactive credentials that could silently satisfy the chain.
• Assign least-privilege Azure roles at the narrowest project, resource, and data scope; review the agent's instance_identity and downstream tool permissions separately from the project client's credential.
• Separate Foundry projects, identities, quotas, networks, and storage across environments and mutually untrusted tenants.
• Do not accept endpoint, subscription, resource group, project, tenant, authority, or credential selection from request data.

Connections and model clients

• Keep connections.get(..., include_credentials=False) and get_default(..., include_credentials=False) at their safe default; enable credential return only inside a narrow adapter and never return, log, trace, cache, or send the resulting credential-bearing Connection to a model.
• Do not return connection secrets, keys, tokens, or full connection objects through APIs, logs, traces, or model context.
• Allowlist model deployments and capabilities before obtaining or using an inference or OpenAI-compatible client.