Cheap beats capable in the attacker's AI toolkit

The model that gets weaponized is not the most capable one, it is the cheapest to run with the weakest filters, and that combination is what makes it the attacker's default. TeamT5's research shows state-affiliated Chinese hacking groups passed over more capable Chinese models like Moonshot's Kimi K3 because Kimi K3 costs too much to run at scale, then routed mundane work and exploit development through DeepSeek, an open-source Chinese model that anyone can self-host and customize. The pattern is documented in scripts and logs TeamT5 obtained in recent months, including from a group called Grimfengxi that used DeepSeek to write exploit code for vulnerabilities. The mechanism is straightforward: a permissive model scales faster than a frontier model with stricter guardrails, even when the frontier model would produce better code, because filtering at scale is a per-request cost, not a per-training cost. Western models from Anthropic and OpenAI sit on the other side, where safety filters force attackers to spend more effort bypassing them, which keeps those models out of the default toolkit. The story is not Chinese versus American AI. It is cost plus permissiveness, not capability, that decides which AI lands in an attacker's hands, and DeepSeek is the current case study.