The gap between open and closed AI models in cyber capabilities keeps shrinking

Open-source AI models can now perform advanced cyberattacks, meaning previously closed security tools may become less effective.

A recent analysis by Vals AI revealed that open-source AI models are rapidly closing the gap in cyber offensive capabilities compared to proprietary models. In simulated cyberattack tests, Deepseek V4.1 Flash, an open-weight model, performed at 72.8% of objectives, closely trailing OpenAI's GPT-6 Sol Daybreak (80.8%) while costing significantly less per run. The harder Lighthouse test saw Deepseek even surpass GPT-6 Sol Daybreak in certain aspects. Longer runtime tests indicated that open models, despite initial slower progress, can achieve substantial results. This trend aligns with a UK's AI Security Institute study showing the gap has shrunk to four to seven months, with open models being substantially cheaper to run and their safety measures proving ineffective. Vals suggests these exploit-capable models can also benefit defenders in identifying vulnerabilities.

The shrinking gap between open and closed AI models in cybersecurity is a critical development, signaling a democratization of advanced offensive capabilities. While proprietary models still hold a slight edge, the speed at which open-weight alternatives are catching up, coupled with their drastically lower cost, suggests a potential shift in the cybersecurity landscape. This also implies that the effectiveness of traditional, closed-source security solutions might diminish as more actors gain access to sophisticated AI-driven attack tools.

This trend necessitates a reevaluation of defensive strategies, as the same AI models that pose threats can also be leveraged for vulnerability detection. Defenders must now consider the rapid pace of open-source development and the potential for these models to be used by a wider range of actors, necessitating more agile and AI-aware security postures. The findings also highlight that current benchmark tests may underestimate the full potential of open models due to their extended progress curves.