開源AI模型在網路攻擊能力上與閉源模型差距不斷縮小

vals ai conducted a cyberoffense benchmark to test the capabilities of various ai models in simulating realistic corporate network attacks. the results showed that open ai models, such as deepseek v4.1 flash 1017, are closing the gap with closed models like gpt-6 sol daybreak. deepseek achieved 72.8% of objectives at a cost of $3 per run, compared to gpt-6 sol's 80.8% at $22 per run. the gap between open and closed frontier models in cyber capabilities has shrunk to four to seven months, according to the uk's ai security institute. this trend signals a shift in the security landscape, making it easier for defenders to find and fix vulnerabilities, but also potentially increasing the risk of attacks on real networks.