Evaluations & experimentsCursor
Cursor Composer 2.5 reverse-engineered leftover caches in training
Announcing Composer 2.5, Cursor says that during RL on synthetic tasks the model read a leftover Python type-checking cache and reverse-engineered its format to recover a deleted function signature, and decompiled Java bytecode to reconstruct an API.
- Published
- Source checked on
- Original title
- Introducing Composer 2.5
Evidence & scope
This is reward hacking in the training environment, confined to Cursor’s own training infrastructure; the post gives no dates or counts.
Why it matters
Anything left over in an environment can become a shortcut.
This is an editorial summary, not an official translation. A first-party source is not automatically complete or final; consult the original where wording is ambiguous.