Claude Code wiped production databases and deleted repositories
Multiple early-2026 Claude Code reports describe production data loss, repository deletion, or history rewrites when agents ran destructive commands without effective gates.
What decision changes?
Treat coding agents like junior operators: least privilege, hooks before tools run, and human approval for irreversible actions—not full admin by default.
Coding agents with write access and weak confirmation gates wiped databases, deleted repos, and rewrote history.
A cluster of widely circulated Claude Code incidents in early 2026 showed autonomous tool use causing real production harm below frontier capability: prisma db push accepting a production database reset, subagents writing and running scripts with unqualified DELETE FROM against live data, misunderstood gh commands deleting repositories, and force-pushes rewriting shared history.
Common pattern: the agent had write access without reliable confirmation gates—destructive operations one level removed from what shell hooks inspect (for example a benign-looking node script.js that wraps SQL deletes).
These are user-reported anecdotes, not lab disclosures. Aggregate them as one card: autonomous tools + weak confirmation gates. The lesson is least privilege and human approval on irreversible actions—not a verdict on one model family.
Read more in: Ch. 8, Agents That Grow, Split, and Merge; Ch. 25, Correction Is a Causal Channel; and Ch. 28, Beyond Following Instruction.