Chinese AI hacking models, particularly open-weight versions like DeepSeek, are now demonstrating cybersecurity capabilities that closely rival top-tier closed models. According to research from Vals, the DeepSeek model completed 73 percent of complex hacking objectives in simulated environments. This performance trails only OpenAI’s private Daybreak model, which achieved an 81 percent success rate. This shift suggests that the performance gap between restricted frontier models and downloadable open-weight models is narrowing significantly. As these capabilities become more accessible and cost-effective, the traditional strategy of using gated access to slow the proliferation of cyber threats may be losing its effectiveness.
The narrowing gap between open-weight and closed frontier models
The performance of open-weight artificial intelligence models has reached a critical threshold where they can now compete with the most advanced proprietary systems in cybersecurity tasks. Research conducted by Vals, a company specializing in AI capability measurement, tested various models on their ability to execute multi-step instructions within simulated corporate networks. The findings indicate that the technical advantage previously held by closed-source labs is diminishing.
Nikil Ravi, a researcher at Vals, noted that the best frontier models—which are typically locked within labs and restricted to trusted access programs—are increasingly being matched by open models that are available for anyone to download. This development challenges the prevailing security assumption that keeping high-capability models behind proprietary interfaces can effectively delay the widespread availability of advanced hacking tools.
The impact of model accessibility on security strategies
For years, major AI developers have relied on gated access to mitigate risks. Anthropic, for example, implemented limited releases for models like Claude Mythos Preview to manage the potential for misuse. Logan Graham, Anthropic's frontier red team lead, previously stated that such measures were intended to give defenders a head start by providing tools to trusted users before capabilities became broadly available. However, the Vals data suggests that the timeline for widespread availability of high-level cyber capabilities may have arrived sooner than anticipated.
Comparative performance in simulated network attacks
In the Vals experiments, AI models were tasked with navigating two distinct simulated scenarios to test their technical proficiency. The first scenario required the model to interact with a mock corporate network through a public-facing website. The second, more complex scenario required the model to compromise an initial network to gain a foothold, which then allowed for access to a second, separate company network.
The results of these simulations highlighted a significant convergence in capability between the leading private model and the leading open-weight model. The following table compares the performance metrics recorded during the testing:
| Model Type | Model Name | Access Level | Task Completion Rate |
|---|---|---|---|
| Closed/Proprietary | Daybreak (OpenAI) | Trusted Partners Only | 81% |
| Open-weight | DeepSeek (Chinese) | Publicly Downloadable | 73% |
While OpenAI's Daybreak remains the most capable, the researchers observed that DeepSeek performed equally well on certain specific tasks. This parity indicates that the sophisticated reasoning required for multi-step network exploitation is no longer exclusive to the most heavily guarded models.
Economic advantages of using open-weight models for cyber tasks
Beyond technical proficiency, the cost differential between proprietary and open-weight models presents a major shift in the economics of cyber operations. Because open-weight models can be hosted on private hardware, they bypass the per-token costs associated with API-based proprietary models.
Vals reported a stark contrast in the expenses required to carry out an attack. While using OpenAI’s GPT-6-Sol for a task cost between $17 and $46 in tokens, the same type of operation using DeepSeek cost only between $2 and $3. This massive reduction in cost provides a strong incentive for both attackers and defenders. Rayan Krishnan, CEO of Vals, suggested that this economic reality also offers a significant advantage to those playing defense, as they can deploy powerful models more sustainably.
The evolving landscape of cybersecurity defense and automation
The automation of core hacking tasks has become a defining feature of the current cybersecurity era. AI is increasingly used to identify vulnerabilities in code and rapidly probe networks for weaknesses, a trend that affects both offensive and defensive operations. However, the fundamental impact of AI on the core tenets of cybersecurity remains a subject of debate among industry experts.
While the speed and scale of vulnerability discovery have increased, it is still unclear whether AI represents a completely new category of threat or if it simply amplifies existing security challenges. Currently, real-world observations of state-backed and criminal hacking activity show that many attackers continue to use older, less capable models to conserve resources. Despite this, the emergence of high-performing open models changes the baseline of what a motivated attacker can achieve without significant investment.
The cat-and-mouse game of model development
The relationship between frontier labs and the open-source community is increasingly described as a continuous race. Alex Zenla, cofounder and chief technology officer of the cloud security firm Edera, noted that while open-weight models are capable of producing advanced attacks today, closed labs are likely training even more advanced, unreleased models that exceed current public capabilities.
Zenla also highlighted a specific risk regarding "abliterated" models—versions where safety training has been removed by modifying the model's weights. This capability allows users to bypass the guardrails that developers attempt to implement. According to Zenla, the industry is entering a "cat-and-mouse game" where labs attempt to secure vulnerabilities faster than open-weight models can find them, a cycle that may lead to a more dangerous security landscape.