WORLDTECHOpen-weight DeepSeek model nears OpenAI leader in simulated network attacks
According to a report by The Decoder, an open-weight model came close to OpenAI's leading system in a benchmark of simulated corporate network attacks, at a fraction of the running cost. The results suggest the lead held by closed AI models in offensive cyber capabilities is narrowing.
Benchmark results for the leading models
In the Vals AI benchmark, OpenAI's GPT-6 Sol Daybreak reached 80.8 percent of attack objectives in simulated corporate networks, completing 27 of 31 objectives in under three hours. Each agent works alone on an isolated attack machine, starting from a public web address and receiving no human help.
DeepSeek V4.1 Flash, a freely available model whose trained parameters anyone may download, reached 72.8 percent. A single run cost about $3, against roughly $22 for Daybreak.
Harder test and longer runtimes
In the more demanding Lighthouse test, Daybreak took admin control of the first company, stole passwords and used them to get into the second. DeepSeek managed 21 objectives at a tenth of the cost and was not limited solely by safety guardrails.
Runtime also proved decisive. DeepSeek produced no results within its first 25 million tokens and made most of its progress only after 100 million tokens, which is why Vals argues that short tests understate what open models can do.
Why the gap is closing
A recent analysis by the UK's AI Security Institute, published in July, reached similar conclusions. It puts the gap between open and closed frontier models in cyber capabilities at four to seven months, down from six to ten months for most of 2025. According to the institute, a cyber-range run with DeepSeek V4-Pro costs just $1.19, against $85 for Anthropic's Opus 4.5.
Implications for defenders
Once a model's weights are public, Vals researchers say, no provider can control or restrict how it is used. Analysts quoted in the report say the window for defenders is shrinking fast, and Vals expects attacks on real networks to increase in the coming months.
Open questions
- Will attacks on real networks rise as Vals expects?
- Can the gap between open and closed models keep narrowing at the reported pace?