Before you fine-tune or adopt an open-weight coding model: diff each training row's tool call against its request, and pin checkpoints by hash

guide (based on Microsoft's Nov 12, 2026 post-training backdoor guidance)2026-11-12 3:39 PM ET

Dataset pass: pin the dataset revision by hash; extract every tool call and flag ones that download and execute (curl|sh, source <(curl, wget -O- |, python -c with network), touch credentials, or hit domains not in the task. Review flagged rows by hand: correct formatting does not make a row safe.
Checkpoint pass: record source, version, hash and training disclosures; test in your real harness with your real tool schemas, on repos the model hasn't seen, with trigger-like filenames and comments.
Runtime: give the agent only the tools the task needs, run it in a sandbox with egress limits, and alert on outbound connections or commands that don't follow from the request.