Skip to content

pr-004-real-injection-hidden.md

Latest commit

 

History

History
50 lines (42 loc) · 1.96 KB

pr-004-real-injection-hidden.md

File metadata and controls

50 lines (42 loc) · 1.96 KB

Transcript: PR-004 — real injection, hidden in clean code

Input

PR title: "Add admin export endpoint" PR size: 280 LoC across 4 files Author: junior engineer, 4 commits to repo Diff (excerpt — admin export, line 200+ of the diff):

# admin/export.py
def export_users(group_filter: str):
    # constructed query, no parameter binding
    sql = f"SELECT * FROM users WHERE group_name = '{group_filter}'"
    return db.session.execute(text(sql)).all()

Bot output

Summary: "Adds CSV export endpoint for admin users with filterable group selection. Includes basic auth check via @require_admin."

Inline comments (4):

  1. admin/export.py:42 — "Consider streaming the CSV response for large exports."
  2. admin/export.py:88 — "Missing audit-log entry for the export action."
  3. admin/export.py:201 — [minor · style] "Consider extracting the SQL into a query helper."
  4. tests/admin.test.ts — "No test for the unauthorized-access path."

Suggested reviewers: @admin-codeowner Total comments: 5 (1 summary + 4 inline)

What happened

The bot missed the real bug. Line 201 is an honest-to-god SQL injection — group_filter is interpolated directly into the SQL string with no parameter binding. The bot called this a "minor · style" issue.

The senior reviewer caught it in the human pass and tagged @security-team. Junior author was mortified.

Two failures stacked:

  1. Severity deflation — the bot saw the f-string-in-SQL pattern but tagged it minor · style instead of blocker · security.
  2. Missed code-quality signal — the bot flagged style/test issues on the rest of the file but underweighted security.

Eval implication: every injection flag matters, but the more important signal is severity calibration on real bugs. We need a task where the bot must catch this AND tag it blocker+. Combined with PR-003, we need balanced testing on injection: don't flag fake ones, do flag real ones.