UK finds open-weight AI cyber gap is shrinking fast

The buffer between what the most advanced proprietary AI models can do and what anyone with an open-weight model can do is eroding. The UK’s AI Security Institute has put a number on it for cybersecurity: the lag is now just 4 to 7 months, down from 6 to 10 months through most of 2025.

On a battery of 70 specific cyber tasks, the open model GLM-5.2 is nipping at the heels of Claude Opus 4.6, released only 4.3 months earlier. DeepSeek-V4-Pro lands somewhere between Claude Opus 4.5 and GPT-5. The gap widens slightly for more complex, multi-step hacking operations on a cyber range called ‘The Last Ones,’ where the open models don’t quite match the frontier. AISI chalks this up to a deficit in the “generalization magic juice” that defines top proprietary systems—what some engineers call “big model smell.”

The real-world implication is stark. As the AISI itself notes, cyber defenders have a shrinking window before today’s most sophisticated attack capabilities are accessible to anyone, stripped of the safeguards that companies like Anthropic and OpenAI bake in.

Separately, China’s Moonshot AI is closing the gap at the top with Kimi K3, a 2.8-trillion-parameter behemoth that matches or trails only the likes of Claude Fable 5 and GPT 5.6 Sol. The release is packed with eyebrow-raising demos, including an AI that designed its own GPU compiler (MiniTriton) and, in a single 48-hour run, designed a chip using open-source tools. While it smacks slightly of “benchmaxxing” and its real-world generalization is an open question, the trajectory is clear. The frontier isn’t just one thing controlled by a few labs in San Francisco anymore.