Open-Weight AI Models Near Closed Rivals in Hacking, Trailing 81% to 73%
Updated
Updated · WIRED · Nov 6
Open-Weight AI Models Near Closed Rivals in Hacking, Trailing 81% to 73%
3 articles · Updated · WIRED · Nov 6
Summary
Vals found open-weight AI models can now carry out complex network intrusions nearly as well as restricted frontier systems in simulated corporate environments, marking a key shift in cyber capability diffusion.
OpenAI’s trusted-access Daybreak led the tests by completing 81% of assigned tasks, while the open-weight Chinese model DeepSeek reached 73% and matched it on some tasks.
Cost differences were stark: Vals estimated GPT-6-Sol attack runs at $17 to $46 in tokens versus $2 to $3 for DeepSeek, giving both attackers and defenders incentives to use open models.
The results undercut frontier labs’ argument that gating advanced models would buy time, a strategy Anthropic had framed in April as a temporary head start for defenders.
Researchers and security firms still say the real-world impact is unsettled, because many attackers continue using older models and closed labs may already hold stronger unreleased systems.
If a $3 open-source AI can hack a corporate network, is the era of gated cybersecurity tools officially dead?
With AI executing nation-state-level breaches in under 10 hours, how can human defenders possibly keep up with the automated onslaught?
The 2026 Collapse of Cybersecurity Barriers: Open-Weight AI Models Slash Costs and Unleash Global Offensive Capabilities
Overview
By 2026, breakthroughs like Sparse Mixture-of-Experts and IndexShare attention made open-weight AI models much cheaper and more efficient, helping them quickly gain nearly 30% of global usage. This cost drop enabled even amateurs to launch major cyberattacks, as seen when a 26-year-old used open models to breach South Korean banks. As open models reached technical parity with closed systems, vendors began restricting sensitive features. Meanwhile, the ability to run these models locally let users remove safety guardrails, fueling new threats like autonomous ransomware. These shifts forced defenders to adopt new security strategies and highlighted the urgent need for updated governance and risk management frameworks.