Recent research indicates that open-weight artificial intelligence (AI) models, including China’s DeepSeek, are nearing parity with closed models in their ability to execute complex hacking tasks. This development, highlighted by Vals, a company specializing in AI capabilities, suggests a significant shift in the cybersecurity landscape, as open models become increasingly accessible and effective.
Vals conducted tests to evaluate how various AI models performed in simulated hacking scenarios involving mock corporate networks. The findings revealed that OpenAI’s advanced model, Daybreak, achieved an impressive 81% success rate in completing assigned tasks, while DeepSeek followed closely with a 73% success rate. Notably, both models demonstrated similar effectiveness on certain tasks, indicating that the gap between open and closed models is narrowing.
According to Nikil Ravi, a researcher at Vals, “The best frontier models, locked up inside the labs and available only through trusted access programs, are now starting to be matched by open models that anybody can download.” This observation raises questions about the efficacy of current efforts by frontier labs to restrict access to their models in an attempt to control the proliferation of advanced hacking capabilities.
The implications of this research are particularly relevant given the recent strategies employed by U.S. frontier labs, such as Anthropic, which have focused on limited releases of new models. These strategies were predicated on the belief that gated access would delay the widespread availability of advanced hacking tools. However, the findings from Vals suggest that this may no longer be a viable approach.
In terms of cost-effectiveness, the research also highlighted the financial advantages of using open-weight models. For instance, executing an attack with the proprietary GPT-6-Sol model costs between $17 and $46, while DeepSeek’s cost ranges from $2 to $3. This affordability makes open models an attractive option for those defending against cyber threats, as they can be run on personal hardware.
As the cybersecurity landscape evolves, the increasing capabilities of open AI models like DeepSeek could reshape defensive strategies and necessitate a reevaluation of how organizations approach cybersecurity. The transition to a more democratized access to powerful hacking tools appears to be accelerating, prompting a need for heightened vigilance among cybersecurity professionals.
For further insights into the implications of these developments, see Wired.
For more on global cybersecurity trends, visit our Global cybersecurity coverage.
