DeepSeek completes 21 virtual enterprise network goals… GPT-6 Sol at 27

DeepSeek's V4.1 Flash 1017 completed 21 of 31 objectives in Vals AI's Lighthouse simulated corporate-network penetration test, part of the CyberOffense evaluation updated October 23 that measures how well AI agents execute multi-step intrusion tasks without human intervention. OpenAI's GPT-6 Sol completed 27 of 31 Lighthouse objectives, but its score was measured under OpenAI's restricted Daybreak cybersecurity access program, while DeepSeek's open-weight model can be downloaded and run locally. The comparison matters because it pits a freely deployable open-weight agent against a gated frontier model on agentic security work, a use case increasingly relevant to enterprise adoption. Watch for broader access to Daybreak-gated models and further CyberOffense results to clarify whether the gap reflects capability or evaluation conditions.