Vals finds Chinese AI models close to top US models on simulated network hacking
Overview
Vals, an AI capabilities measurement company, tested AI models on multi-step hacking tasks across two simulated corporate networks.
OpenAI's Daybreak, available only to trusted partners, completed 81 percent of tasks on average, and DeepSeek completed 73 percent, according to Wired's report on the research.
Vals says an attack costs $17 to $46 in tokens with GPT-6-Sol, compared with $2 to $3 with DeepSeek.
Written by AI from the articles below · updated Nov 6, 2:35 PM ET
Check the sources:
Article timeline
The articles in this story. Times are ET.
- Wired · AINewsFree Chinese AI Models Are Getting Scary Good at Hacking
AIVals, an AI capabilities measurement company, tested models on multi-step hacking tasks across two simulated corporate networks. OpenAI's Daybreak, available only to trusted partners, completed 81 percent of tasks on average, while DeepSeek completed 73 percent. Vals says an attack costs $17 to $46 in tokens with GPT-6-Sol compared with $2 to $3 with DeepSeek.
Heat Trend
Not enough continuous observations to show a trend yet.