DeepSeek completed 73% of simulated corporate-network intrusion tasks in a Vals benchmark. OpenAI’s restricted Daybreak model completed 81%. The tasks included gaining initial access and moving through a network.
The cost gap caught my attention. In Vals’ cost comparison, DeepSeek used roughly $2–$3 in tokens per simulated attack, compared with $17–$46 for the proprietary OpenAI model tested.
These figures come from a controlled lab. What stands out to me is that a downloadable model can already handle many of the same multi-step tasks at a much lower token cost.
If that makes it cheaper for attackers to test more paths into a network, SOC teams will need to get better at connecting the early signs of an intrusion, before it turns into lateral movement.
👀 Will keep an eye on this, clearly China has clear objective with AI. They are trying to make it as useful as possible, so they can integrate into economy. Cybersecurity is quick win for them.