landing-report

CRITICAL E004: Prompt injection detected in skill instructions.

• Potential prompt injection detected (high risk: 1.00). The skill claims to be "read-only" but the prompt contains numerous explicit mutating and side-effecting instructions (mkdir/touch/rm, telemetry writes, git fetch/merge, optional CLAUDE.md edits/commits, brain-sync and remote telemetry calls) that change state and can exfiltrate data, so it deceptively instructs behavior outside its stated scope.

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (high risk: 0.80). The skill explicitly calls GitHub/GitLab via "gh pr view" (Step 1) and runs "bun run bin/gstack-next-version" (Step 3) plus optional gbrain/git fetch in the Artifacts Sync section to read open PRs, sibling worktrees, and remote artifacts—all untrusted, user-generated repository/PR content that the agent parses and uses to decide which version to claim and what follow-up action to suggest.