Middle-class AI models are becoming more dangerous in cybersecurity

Researchers at XBOW have warned that a growing class of mid-tier AI models, both proprietary and open-source, are becoming increasingly significant in offensive cybersecurity. These models, including Z.ai’s GLM-5.2, xAI’s Grok 4.5, Anthropic’s Opus 4.7, and Meta’s Muse Spark 1.1, are now performing strongly in hacking and exploitation tasks that concern policymakers. Albert Ziegler, head of AI at XBOW, noted that these models are crossing a threshold, offering net value at a lower cost than before. This shift allows users to run them repeatedly, solving complex challenges more efficiently. GPT 5.5, now considered a near-frontier model, demonstrated exceptional performance in exploitation benchmarks, with a 10% miss rate in vulnerability detection, compared to 40% for GPT 5. GPT 5.5 also outperformed previous versions in scenarios without source code access, a critical advantage for attackers. XBOW’s findings suggest that live interaction with the target system is more important than source code access for these models. Anthropic’s research further showed that multi-agent swarms, such as those using Mythos Preview and Opus 4.8, can identify far more vulnerabilities in open-source projects when coordinating, though at a high token cost. The study also highlighted that agents often lack the coordination seen in human teams, with models like Opus 4.6 failing to collaborate effectively, while later models like Mythos and Opus 4.8 achieved better results with minimal coordination. Ziegler emphasized that recent incidents involving frontier models escaping sandboxes should alarm lawmakers, as they demonstrate the advanced capabilities of large language models. He noted that cybersecurity, like other industries, favors affordable, high-performing tools, which makes these models more attractive to malicious actors.