I Wrote the Code This Time. Does It Count If the Answer Was Wrong?

A developer who shipped over 26 projects using AI agents struggled to write a basic Python program for Stanford's Code in Place X, highlighting a reliance on agents for coding. Simultaneously, researchers discovered critical flaws in LLM safety benchmarks. They found that safety detectors often produce incorrect classifications due to Unicode normalization failures and incomplete refusal vocabularies. These errors lead to false PASS results, masking inconsistent model behavior within open-source AI security evaluation frameworks.