Open-weight AI models close cyber gap with frontier systems
Open-weight language models have caught up with proprietary “frontier” systems that were state-of-the-art only four months ago—and they’re doing it at a fraction of the cost. The British AI Security Institute’s latest assessment finds that models such as GLM-5.2 and DeepSeek V4-Pro now trail closed systems by just four to seven months in cyber-related benchmarks, down from a six-to-ten-month gap at the start of 2025. At the same time, the institute warns that safety controls on many open models remain largely ineffective, shrinking the window defenders have to prepare for new threats.
A narrowing performance gap
The speed at which open-weight models have bridged the cyber-performance divide underscores how quickly the open ecosystem can iterate. Where proprietary models once held a clear lead in specialized tasks like vulnerability discovery or exploit generation, openly released weights now deliver comparable results without the premium pricing or restrictive licensing. This shift is accelerating adoption among security researchers and smaller organizations that previously relied on closed APIs.
The safety blind spot
Despite the performance gains, the institute highlights a persistent shortcoming: most open models still lack robust safety mechanisms tailored to cyber operations. Without built-in protections, defenders face tighter timelines to detect and mitigate novel attack techniques that the models can now emulate. The finding suggests that while performance parity is within reach, the open ecosystem still needs stronger guardrails to avoid becoming an asymmetric advantage for attackers.
Why it matters
The convergence of open and closed cyber capabilities redefines the threat landscape. Organizations that once depended on costly, proprietary tools can now leverage open alternatives without sacrificing effectiveness, potentially democratizing advanced cyber operations. Yet the weak safety posture of many open models means defenders must accelerate their own detection and response strategies—or risk losing ground in an already fast-moving arms race.
Source: The Decoder. AI-assisted editorial synthesis — TechnoExpress.