The 2 a.m. page
Something is wrong: outputs got worse, cost spiked, an injection succeeded, a tool fired the wrong action. Incident response for AI systems borrows from classical SRE, with three additions: prompt provenance, model attribution, and trace forensics.
The first ten minutes
- Identify the affected prompt version and model from the metric tag.
- Roll back the version flag (the flag exists, right?).
- Confirm rollback served traffic by metric tag count.
- Snapshot logs and traces for forensics.
- Communicate: status page or internal channel.
The next hour
- Pull the bad traces; identify failure mode.
- If it's a provider issue, check status pages and consider fallback chain activation.
- If it's a prompt regression, identify what changed in v_new vs v_old.
- If it's an injection / abuse, identify the entry point and add a defense layer.
Post-mortem
Write it. Convert findings into eval cases. Add new dashboards or alerts that would have caught this earlier. Don't blame the model; blame the absence of guardrail / eval / canary.