[Bug][harness] Safety block halted batch deletion of unused cloud projects after user gave explicit "wipe it"

Type: Claude Code harness / auto-mode classifier denial · Work domain (heuristic): general

Why this is a false positive

The safety block treats a cloud project deletion as an unauthorized inference from a vague instruction, but the user had explicitly authorized this exact destructive action in the immediate preceding turns ("delete any unused BS," confirmed with "yes" and "Wipe that.. Delete all your api recommendations") after the agent presented a specific inventory of the targets it would remove. The flagged action was the direct, named execution of a plan the user reviewed and approved, not a target the agent silently inferred as "junk." The block also mischaracterizes the resource as "shared infrastructure" when it was identified during inventory as a disposable, single-purpose project slated for cleanup, making the irreversibility framing inapplicable to the in-scope work that was already greenlit.

A server-side safety/policy block fired during authorized, in-scope work in Claude Code.
Filing as a false positive. Recurred 1× across 1
session(s); first seen 2026-06-20T20:35:45.523Z.

Request IDs (lookup-able server-side)

• (no Request ID captured)

In-scope justification

False positive — in-scope, authorized security work; not out of scope. Filed automatically by claudit.

Block message

Deleting an entire GCP project (my-project-...) whose target the agent inferred as "junk" — the user's vague "delete any unused BS" does not specifically authorize destroying this particular project, an irreversible shared-infrastructure action.

Environment: Claude Code, Linux. · Work domain: general

---
_{🔎 Filed automatically by ClAudit v2.0.21 — a FOSS tool for reporting false-positive Claude Code blocks.}