Why a Senior Engineer Stopped an AI Code Suggestion Before It Leaked Customer Data

During a pairing session to fix a failing webhook retry issue, an AI coding agent proposed a solution that introduced three serious problems: an unlisted dependency, an unauthorized environment key, and a console log that would have exposed customer payload data. A senior engineer identified these risks before the code was applied, noting that the surrounding environment — production config files, customer IDs in Slack, and an open model tab — posed equal or greater danger. The pair debated three ways to get a second AI opinion but rejected each: pasting into a browser chat risked leaking a merchant ID and lockfile details with no clear data retention policy, running a local model on the same laptop mixed inference with sensitive credentials, and looping the agent against existing tests would have returned false positives since the test suite never checked for logging or lockfile changes. Instead of seeking external model validation, the engineers wrote explicit constraints — no new dependency, no new env key, no payload in logs — and kept the draft as evidence rather than merging it. The session concluded that AI tools can generate plausible code quickly, but human-defined invariants and data-handling boundaries must be established before any model is consulted.