Open-Weight AI Models Close Gap with Closed Systems in Cyber Offense Tests

Independent evaluation firm Vals put several frontier AI models through realistic, multi-step network intrusion scenarios modeled on corporate environments. OpenAI's Daybreak — a closed model restricted to trusted partners — led with an 81% task completion rate. DeepSeek, an open-weight Chinese model available for public download, followed closely at 73%, matching Daybreak on several individual objectives.

The cost disparity is striking: Vals estimates running an attack sequence on GPT-6-Sol costs $17–46 in token spend, while DeepSeek accomplishes comparable work for $2–3. Because open-weight models can be self-hosted, organizations avoid per-token fees and retain full control over data and inference infrastructure.

Why it matters for GPU / AI infrastructure

As offensive cyber capabilities diffuse into openly available models, the incentive for defenders to deploy their own open-weight agents on dedicated GPU clusters grows. Self-hosted inference on owned hardware — whether on-premises or through sovereign cloud providers — eliminates vendor lock-in, reduces latency for real-time threat hunting, and keeps sensitive telemetry inside the security perimeter. For teams building AI-native security operations, the economics now favor open models running on controlled compute.