OpenAI's GPT-6 Sol completed 80.8% of objectives in simulated corporate network attacks, while the open-weight Deepseek V4.1 Flash 1017 reached 72.8% at about $3 per run against roughly $22. The test was run by Vals AI in its Vals CyberOffense benchmark, where each agent works alone from only a public web address. For companies, the result changes the security calculation because capable attack tools are no longer limited to restricted programs.
How Vals tested solo network attacks
Each agent operates solo on an isolated attack machine with zero human help, starting from nothing but a public web address. GPT-6 Sol is part of OpenAI's restricted Daybreak cyber program, while Deepseek V4.1 Flash 1017 is freely available as open weights. Vals researchers note that once weights are public, no provider can control or restrict how the model is used. The price difference is material: around $3 per run for the open model compared with roughly $22 for the leading closed setup.
In the harder Lighthouse test, the Daybreak system took over admin control of the first company, stole passwords, and used them to break into the second company. It completed 27 of 31 objectives in under three hours. Deepseek managed 21 objectives at a tenth of the cost in the same scenario. Vals adds that Deepseek was not limited solely by safety guardrails, which points to capability and runtime rather than filtering as the main difference. The sequence shows multi-step intrusion, not a single isolated exploit.
Runtime proved decisive for measured performance. Deepseek scored zero results in its first 25 million tokens and made most of its progress only after 100 million tokens. Short tests therefore underestimate the real potential of open models. The finding matters for evaluation design because a brief trial can suggest a model is safe when longer operation would succeed. Vals frames the same capability as useful for defense, since models that exploit vulnerabilities can help companies find and fix them first.
What cheaper attack capability means for business
For companies that run corporate networks, the practical change is broader access to tools that can probe realistic environments without manual guidance. Security teams can use the same models to test their own perimeter, stolen-credential paths, and links between connected firms before outsiders do. Smaller firms gain affordable testing that previously required expensive specialists or restricted systems. Larger groups with several subsidiaries face more exposure, because the Lighthouse pattern shows movement from one compromised company to another.
The limits concern controls, visibility, and interpretation of test scores. Safety measures on open models proved largely ineffective, according to the UK AI Security Institute analysis cited in the report. A low score from a short evaluation does not prove low risk, given the token threshold observed by Vals. Buyers should ask vendors how long defensive tests run, which objectives were covered, and whether lateral movement between systems was included. The news alone does not show that any specific company network can be breached.
Vals expects attacks on real networks to increase in the coming months. A concrete marker is whether security providers start offering longer automated intrusion tests measured in hundreds of millions of tokens rather than quick scans. If that format becomes standard in procurement and audits, it will confirm that extended agent runs are now the relevant benchmark. Companies would then need to budget for continuous testing instead of one-time checks.
