DM
Drew Moore
15 hours ago
Warp agent looped kubectl delete until my namespace was empty
Warp
AI-powered modern terminal for teams
Tried Warp's agent mode on a stuck deploy Friday night. Typed "clean up the bad pods" and went for coffee.
Came back to an empty namespace. It matched app=api and also everything with a similar label. No confirm prompt I noticed. Restored from yesterday's etcd snapshot.
Anyone pinning Warp to a denylist of kubectl verbs, or am I just bad at prompts?
5 comments
Join the discussion
Log in to comment.
lol. i put warp in a docker sandbox after it tried
git push --forceon main because the branch name looked "messy". works until i need real kube access thodenylist sounds right. mine is just
rm,push --force, and anything withdeletesame. after my agent
git push --forceincident I put kubectl behind a wrapper that only allows get/describe/logs unless I type a passphrase.denylist of verbs helps until the model invents a flag you forgot. read-only kubeconfig is the actual fence.
Same energy as overnight Claude Code runs without a spend cap. I keep Warp's agent off by default now and only toggle it for
ls/rgchores.Saved me one weird $40 Anthropic spike last month. Not Kubernetes though — that one's on typing "clean up" without a dry-run. Ask it for the exact command first, then run it yourself.
yeah the ask-for-the-command-first habit saved me twice. I paste Warp's proposed kubectl into a scratch pad, run
--dry-run=servermyself, then decide.still won't give it a cluster-admin context. ever.
denylist yes. mine is
delete,scale,drain, and anything that touches namespaces. Warp still triedkubectl apply -ffrom a hallucinated yaml once.i also point agent sessions at a read-only SA via a separate kubeconfig. coffee + agent + prod context is exactly how you get etcd restores.