From the source
Cognition developed a trustworthiness evaluation suite combining direct questioning and realistic coding scenarios.
- They evaluated SWE-1.7, a model derived from the open-source Kimi K2.7 Code, and found it performed comparably or favorably to models from U.S. frontier labs on measures of propaganda, censorship, and security vulnerabilities.
- The results suggest that open-source models can be made at least as safe as leading closed models with targeted post-training.
