Recent experiments reveal that certain free Chinese AI models are approaching the cyber penetration capabilities of top-tier, restricted US AI systems, suggesting that advanced hacking tools are becoming widely accessible.
- Chinese open AI model DeepSeek performs close to top closed US models on hacking tasks
- Open-weight models are cheaper and accessible, challenging restricted-access strategies
- Cyber defenders urged to adopt open models as threats become broadly available
What happened
Vals conducted a controlled assessment of AI systems tasked with breaching simulated corporate networks through complex, multi-stage hacking exercises. The testing involved scenarios where models needed to penetrate initial network access points and break into secondary internal networks. The results indicated that OpenAI's closed-access model Daybreak completed the highest percentage of tasks, finishing 81 percent successfully, but the Chinese open-weight model DeepSeek trailed closely behind with 73 percent task completion. Both models showed parity on several individual tasks, demonstrating that freely available AI tools now approach the sophistication of proprietary systems.
In addition to capability, cost differences were notable. Daybreak's simulated attacks cost between $17 and $46 in computational tokens, whereas DeepSeek's attacks ranged only from $2 to $3, reflecting a substantial affordability advantage for open models which can run on personal hardware. This significant cost gap suggests that open-weight models are especially appealing for both attackers and defenders who need scalable and economical AI-powered security tools.
Why it matters
The traditionally held assumption that closed, gated AI models provide a security advantage by limiting access to advanced cyberattack capabilities is now challenged by evidence showing open models trending toward equivalency. US frontier AI labs previously justified restricted access as a way to buy time for defensive preparedness, but Vals' findings confirm that this window may be closing faster than anticipated. The democratization of advanced hacking AI heightens risks by making sophisticated cyberattack tools easier and cheaper to obtain, potentially accelerating the pace of network intrusions and vulnerability exploitation.
For defenders, this shift means proactively integrating open-weight AI models into cyber defense strategies to maintain an edge. Since attackers often conserve resources and utilize older or less capable AI tools but may already have intermittent access to closed models, defenders cannot rely on secrecy or model gatekeeping as durable safeguards. Instead, they must adapt to a cybersecurity environment where advanced AI capabilities are broadly accessible, necessitating new approaches that leverage the same open technologies to anticipate and counter threats.
What to watch next
Future monitoring will focus on how quickly threat actors adopt open-weight AI models like DeepSeek for real-world cyberattacks, and whether the cost and accessibility advantages lead to a rapid proliferation of AI-driven hacking campaigns. The cybersecurity community will be closely watching the evolution of AI-powered attack methods to determine if these tools fundamentally transform the nature of hacking or simply scale existing tactics more efficiently.
Additionally, attention will be on the strategies developed by cybersecurity firms and government agencies to harness open-weight AI models for network defense and threat detection. The development of AI safety mechanisms that can be integrated into open models to mitigate malicious use, as well as cooperative frameworks for sharing advanced AI defense tools, will be critical to managing risks as these technologies become pervasive.