The gap between open and closed AI models in cyber capabilities keeps shrinking
via The Decoder · 32 views · 1w ago
AI's cyber attack capabilities are getting scary fast. A new benchmark, Vals AI, reveals OpenAI's GPT-6 Sol Daybreak hitting over 80% of attack objectives in simulated corporate networks. This shows rapid advancement in sophisticated cyber offensive capability from closed AI models. The kicker? The open-weight model DeepSeek V4.1 Flash wasn't far behind. It reached nearly 73% of those same attack objectives. That's a remarkably narrow gap, just eight points from the top-tier closed model. Even more concerning is the cost. DeepSeek achieved this impressive feat at only a tenth of the expense. This dramatic cost reduction puts highly advanced cyber tools within reach of many more actors. The implications are clear: the shrinking gap between open and closed AI models means sophisticated cyber attacks are becoming cheaper and easier to execute. This rapid democratization of powerful offensive AI presents serious new challenges for cybersecurity defenses.
Read the original story →Loading comments…
Log in to join the conversation.