DeepSeek 4.2 Flash outperforms frontier models for offensive security

XBOW evaluations of AI models for offensive security reveal that the most expensive and tightly controlled frontier models are not always the most effective or economical choice. Researchers found that while newest frontier models perform well at higher budgets, open-weight alternatives like DeepSeek 4.2 Flash outperform them at lower budgets, challenging assumptions that access control alone can prevent misuse. Open article →