Chinese AI models match OpenAI's hacking skills
open-weight ai models from china, like deepseek and glm, are almost as good at hacking as openai's best cyber model, daybreak, which is only available to trusted partners.
vals tested ai models in two simulated scenarios, giving them specific tasks to achieve. openai's daybreak model, available only to trusted partners, completed 81% of tasks in an average of two tests. deepseek, a chinese open-weight model, completed 73% of objectives on average. the research also showed that using open-weight models is much cheaper, with costs ranging from $2 to $3 for deepseek compared to $17 to $46 for gpt-6-sol. this could provide an incentive for those playing defense to use open models rather than proprietary ones.