The gap between open and closed AI models in cyber capabilities keeps shrinking
What happened
OpenAI’s GPT-6 Sol Daybreak and the open-weight model DeepSeek V4.1 Flash have closed the performance gap in cyberattack simulations. GPT-6 achieved 80.8 percent success in hitting attack objectives in the Vals AI benchmark, a widely recognized test using simulated corporate networks. DeepSeek V4.1 Flash performed at 72.8 percent, delivering nearly comparable results at roughly one-tenth of the cost to run.
Why it matters
The narrowing gap between closed and open AI models means powerful cyber capabilities are no longer confined to proprietary platforms. Open models have become too capable to control effectively, shrinking the time defenders have to adapt before attackers deploy these tools. The cost advantage of open models also lowers the barrier for malicious actors to use advanced AI in real-world attacks, increasing the pressure on corporate security teams and incident responders.
This shift weakens the traditional security-by-obscurity advantage held by vendors who restrict access to their models. Defenders can no longer rely on closed models to maintain a lead in AI-driven cyber offense and defense. The quick rise of open-weight models compresses threat windows and requires faster, more automated defense strategies that do not depend on model provenance.
What to watch next
Watch how cybersecurity firms and enterprises integrate AI benchmarks like Vals to test their defenses against both open and closed AI attack models. The market will need new detection tools that can spot AI-driven attack patterns regardless of the model source. Investors and operators should track innovations in cost-effective AI defenses that can scale to address the rising threat from increasingly accessible and capable open models.
Expect AI governance and regulatory attention to sharpen around controlling misuse of open-weight models. However, technical controls alone will not be enough to slow attackers deploying AI-powered cyber tools that have near-parity with closed systems.
AI Quick Briefs Editorial Desk