Physix Frontier · News Briefing Card (Wired · Nov 7, 2026)

Study: Open-source AI models near closed-source frontier in cyberattacks

KEY FACTS

Vals tested multiple AI models' cyberattack capabilities against a simulated corporate network.
OpenAI's closed-source model Daybreak completed an average of 81% of tasks, while DeepSeek completed 73%.
On some tasks, DeepSeek scored the same as Daybreak.
DeepSeek's token cost to execute attacks was $2 to $3, versus $17 to $46 for GPT-6-Sol.
Vals said closed-source models no longer hold an absolute advantage in cyber skills.

KEY DATA

81%Daybreak task completion rate

73%DeepSeek task completion rate

$2-3DeepSeek attack token cost

$17-46GPT-6-Sol attack token cost

PHYSIX OBSERVATION

Open-source models matching closed-source cyberattack capability means the strategy of delaying risk proliferation by restricting model access has largely failed. More alarming is the cost gap: attackers can reproduce closed-source capability with open-source models for a few dollars. Defenders should quickly put open-source models to work on vulnerability hunting rather than continuing to bet on access barriers.

Source: Wired report