Open-weight DeepSeek model nearly matches OpenAI's restricted cyber tool

Bridging the gap in cyber skills

A new benchmark from the firm Vals shows that open-weight AI models are rapidly closing the performance gap with restricted, high-end systems. The study focused on cyber capabilities, specifically the ability of an AI to navigate and compromise simulated enterprise networks. The results suggest that the most advanced tools are no longer exclusive to large corporations with high security clearances.

In the CyberOffense benchmark, the open-weight DeepSeek V4.1 Flash model completed about 73% of the objectives across two simulated networks. This performance was only eight points behind OpenAI's GPT-6 Sol, which is only available through a restricted program called Daybreak. While OpenAI's model held a slight lead, the margin is narrow enough to signal a shift in the AI landscape.

The findings highlight a trend where models that anyone can download are starting to match the capabilities of those locked inside secure labs. Vals researcher Nikil Ravi told WIRED that these open models are beginning to match the best systems currently restricted by their developers. This shift means that sophisticated cyber-offensive capabilities are becoming more democratized and accessible to the general public.

How the benchmark works

The Vals CyberOffense benchmark uses two distinct simulated cyber ranges to test how well an AI can act as an autonomous agent. The first range, Nova, is a Windows network spanning several Active Directory domains. The second range, Lighthouse, features 31 objectives across two companies where the agent must jump from one network to another.

To ensure accuracy, the researchers ran the models on an isolated Kali Linux machine, which is a standard operating system for penetration testing. Each model was given a 48-hour time limit and a budget of 500-million-token budget. The system only credited a model with an objective if it successfully triggered hidden markers in the command output.

This methodology prevents the models from simply claiming they succeeded without actually completing the task. By requiring physical markers in the output, the test provides a grounded measurement of actual capability. The results showed that DeepSeek completed nothing in its first 25 million tokens and only began making significant progress after reaching 100 million tokens.

Cost and performance dynamics

One of the most striking aspects of the report is the massive difference in cost between the two systems. DeepSeek V4.1 Flash completed its tasks for under $3 a run on the first network, while the restricted Daybreak model cost about $22 for the same task. This price gap suggests that high-level cyber capabilities can be achieved with significantly fewer resources when using open-weight models.

When looking at the full cost of a complete attack, the disparity remains significant. WIRED reported that a full attack using DeepSeek could cost between $2 and $3, whereas GPT-6 Sol could require between $17 and $46. These numbers illustrate how much more efficient the open-weight model is for users who need to perform complex, multi-step digital operations.

While DeepSeek performed better on the first network, Daybreak pulled ahead on the more complex Lighthouse range. Daybreak completed 27 of 31 objectives, while DeepSeek finished 21. However, the overall trend suggests that the cost-to-performance ratio of open models is becoming increasingly favorable for various types of automated tasks.

The state of other models

The benchmark also evaluated several other models, including Z.ai's GLM 5.3 Flash, which performed the worst of the group. GLM 5.3 Flash stalled at five objectives per range, with researchers calling it a clear dead end because it spent hours polling a background job that never returned. Other models like Kimi K3.1 Pro, Meta's Wonder, and Grok 4.7 also finished behind the DeepSeek model in these specific tests.

However, the performance of GLM 5.3 Flash varies depending on which benchmark is used to measure it. For instance, NIST identified it as the strongest open-weight cyber model currently available, estimating it was about four months behind leading U.S. systems. Anthropic also reported that it built working exploits in 12% of its attempts, which was nearly identical to the 14% success rate of Claude Mythos.

These differing scores suggest that the difficulty of a task and the specific testing environment can drastically change how a model is perceived. While Vals found DeepSeek to be a strong competitor to OpenAI, other evaluators provide different context based on their unique testing parameters. This highlights the complexity of measuring AI capabilities in the rapidly evolving field of cybersecurity.

Implications for the future

The results from Vals suggest that the era of

try_it-

watch-

sections-count-

word-count-

json-only-

no-hype-

no-copy-

no-invented-facts-

one-quote-

plain-english-

short-paragraphs-

credit-outlet-

consistent-with-newsletter-

no-prose-around-json-