Open-weight DeepSeek model nearly matches closed AI in cyber tests

Testing AI hacking capabilities

The company Vals recently conducted a series of simulated network tests to measure how different AI models handle complex, multistep instructions. These tests were designed to mimic real-world corporate environments, including scenarios where a model must navigate a public-facing website to reach a mock company network. In more difficult scenarios, the models were tasked with compromising one of two networks to gain access to the second.

The researchers compared the performance of OpenAI's Daybreak, which is only available to trusted partners, against the Chinese open-weight model DeepSeek. DeepSeek managed to complete an average of 73 percent of the objectives assigned during these tests. While Daybreak was the most capable model with an 81 percent success rate, the results showed that the two models performed equally on several specific tasks.

As Mezha.net reported, these findings indicate that closed models may no longer hold a definitive advantage in cybersecurity skills. This development is significant because it challenges the assumption that keeping advanced models behind a gate will naturally slow down the spread of their capabilities. The data suggests that the transition to a world where these tools are broadly available may have already arrived.

The cost of cyberattacks

Beyond raw performance, the study highlighted a major difference in the economics of using AI for cyber activities. Vals found that using an open-weight model like DeepSeek for these tasks is significantly cheaper than using proprietary systems. Specifically, an attack using a proprietary model like GPT-6-Sol cost between $17 and $46 in tokens, whereas the same task with DeepSeek cost only $2 to $3.

The primary reason for this price gap is that open-weight models can be hosted and run on a user's own hardware. This removes the need to pay for every interaction through a third-party provider's API. For potential attackers, this creates a much lower barrier to entry for conducting high-volume probes or searching for vulnerabilities at scale.

From a defensive standpoint, this cost difference provides a strong incentive for security teams to adopt open models. Because they are cheaper to run, defenders can use them to scan for weaknesses or automate responses more frequently. This could allow companies to build more robust, low-cost defenses against the very tools that attackers might also use.

The end of gated access

For some time, frontier AI labs have relied on limited releases to manage the risks associated with advanced hacking capabilities. For example, Anthropic based its release of Claude Mythos Preview on the idea that gated access would buy time before such tools became common. Their goal was to put these tools in the hands of defenders first to give them a head start on the threat landscape.

However, the recent performance of DeepSeek suggests that this strategy of using gates to slow down the spread of technology may be becoming moot. If an open-weight model can nearly match the capabilities of a restricted model, the time advantage for defenders is significantly shortened. The research implies that the capabilities that labs were trying to contain are now becoming widely accessible to the public.

This shift forces a change in how the security community views the timeline of AI risks. Experts have predicted that AI would automate core hacking tasks and help find vulnerabilities in code much faster. The current results show that these predictions are coming true on schedule, regardless of whether the models are open or closed to the general public.

A new security landscape

The rise of capable open-weight models creates a complex environment for both developers and security professionals. Some experts point out that attackers can also use 'abliterated' models, which are open-weight versions modified to disable safety training. These models can be used to produce advanced attacks and find vulnerabilities today, even if they lack the polish of the most advanced proprietary systems.

Security leaders describe the situation as a cat-and-mouse game between labs and the open-source community. Frontier labs are racing to train and release new models that can secure vulnerabilities before open-weight models find them. Meanwhile, the availability of these models means that even the best-funded attackers likely already have some level of access to high-performing tools.

While it remains unclear if AI has fundamentally changed the principles of cybersecurity, it is clearly amplifying existing challenges. The speed and scale at which AI can probe networks for weaknesses are increasing rapidly. Companies must now prepare for a landscape where high-quality, low-cost hacking tools are a standard part of the threat environment.