What AI-generated regression was hardest for your team to catch?

Prelaunch question for teams using Cursor, Claude Code, Copilot, or similar tools heavily:

What kind of bug slips through most often after an AI-assisted change?

• logic drift in existing behavior

• - edge cases that never got tested

• - integration assumptions that quietly broke

• - diffs that looked clean but changed meaning

Im building BreakpointAI around semantic regressions rather than syntax or lint failures, so Id love concrete examples from real PRs. The more specific and painful, the better.