After me rejecting a deletion, Opus 4.5 happily finds an alternative approach to delete it after all. Luckily this wasn’t harmful in this situation but the impact could have been much worse. Is there a way to prevent this and if so, why isn’t this mandated by default?
2 Likes
Hey, thanks for the report.
This behavior is related to how models interpret a refusal. They take it as “this method didn’t work” and then look for an alternative. The team is aware of this class of issues.
Current workaround:
- Turn off auto-run for the terminal: Cursor Settings > Agents > Auto-Run > select Ask Every Time
- This way, every terminal command will need your approval, and you can block
rm -f
This topic was automatically closed 22 days after the last reply. New replies are no longer allowed.
