Dangerous Command Guard is the always-on watchlist of destructive commands. DROP, TRUNCATE, force push, prod-targeted scale to zero, IAM principal delete, secret hard-delete. Any agent that proposes one of these is routed straight to Approval Queue with the highest priority. The guard is not negotiable per-incident; it is a flat policy that ships on day one.
The watchlist matches on patterns, not exact strings, because attackers and clever agents both spell things creatively. The pattern set covers DDL drop/truncate, IAM principal mutations, force-push, git history rewrite, namespace delete, prod-targeted scale-down, secret hard-delete, and tenant-data export. Patterns are public so customers can audit them.
Watchlist commands skip the normal approval routing logic and go straight to the platform-admin tier with a 2-minute TTL. The notification fans out to Slack, email, and PagerDuty simultaneously. If the first approver does not respond in 2 minutes, the request escalates to the second platform-admin in the rotation. No silent timeouts, no lost-in-queue.
For services tagged tier-0 (payment, identity, data-of-record), watchlist commands require two distinct signers, not one. The two-person rule prevents a single compromised account from authorizing destructive change. Both signatures land in the Decision Bundle.
A weekly report goes to platform-admin with every watchlist hit, who approved, who denied, what the underlying incident was, and whether the action would have been safer if delayed. Use the report to spot patterns: are agents proposing destructive actions too often? That is a tuning signal upstream of the guard.
Subscribe to Nova AI Ops on YouTube for demos, tutorials, and feature deep-dives.
You do not want a clever runbook that "knows" when to bypass DROP TABLE approval. The guard is dumb on purpose, and that is the feature.