Free Chinese AI Models Are Getting Scary Good at Hacking

The cost to perform AI-driven cyberattacks has fallen significantly for those using open-source models.

A recent test by AI capability measurement company Vals revealed that free, open-source AI models are now matching the advanced hacking capabilities of proprietary, closed-door models. Vals simulated complex network intrusion tasks on mock corporate environments, finding that the Chinese open-weight model DeepSeek performed nearly as well as OpenAI's most advanced closed model, Daybreak. DeepSeek completed 73% of assigned tasks compared to Daybreak's 81%, with some tasks scoring equally. Furthermore, using open-weight models like DeepSeek for such tasks is significantly cheaper, costing an estimated $2-$3 in tokens versus $17-$46 for GPT-6-Sol. This development challenges the strategy of frontier labs like Anthropic, which limited access to their models based on the belief that gated access would delay the proliferation of advanced cyber capabilities.

The diminishing performance gap between closed and open AI models in cybersecurity tasks suggests that efforts to restrict access to cutting-edge AI for security reasons may be largely ineffective. This leveling means that the advanced offensive capabilities previously thought to be exclusive to well-funded research labs are now accessible to a much broader range of actors. Consequently, the defensive strategies based on maintaining a technological advantage through exclusivity are becoming obsolete, forcing a reassessment of how to prepare for a landscape where sophisticated tools are widely available.

The economic advantage of open-weight models for hacking is a critical factor, as their lower operational costs make them an attractive option for attackers seeking to maximize their resources. This presents a dilemma for defenders, who must now consider how to counter threats that are not only powerful but also economically viable for a wider array of malicious actors. The rapid democratization of these advanced cyber capabilities signals a potential acceleration in the discovery and exploitation of vulnerabilities, underscoring the urgent need for more robust and scalable defensive measures.