New model from Moonshot AI is ‘extremely close’ to OpenAI’s flagship after ‘very big jump’ in capabilities, researcher says Kimi K3 modelrivals top US systems in spotting cybersecurity flaws, according to early research, stoking fears in Washington that strict safety guard rails are putting American
artificial intelligencefirms at a competitive disadvantage. OpenAI’s flagship GPT-5.6 Sol at a fraction of the cost, Swiss firm Aikido Security said in a report on Sunday.
Aikido said it tested leading AI models against 26 known vulnerabilities to evaluate bug-detection rates and cost-efficiency. Kimi K3 discovered 23 of the vulnerabilities, matching the performance of OpenAI’s mid-tier GPT-5.6 Terra, while running at a fourth of the cost of the flagship Sol model.
Because the test was private and used recently discovered vulnerabilities, Kimi K3 “was not able to train on them”, Aikido researcher Philippe Dourassov said on social media. “The performances reflect a very big jump in Kimi models’ capabilities ... Open-source models are no longer behind.”
[Zhipu AI](https://www.scmp.com/topics/beijing-zhipu-huazhang-technology?module=inline&pgtype=article)’s GLM-5.2, which was
[launched last month](https://www.scmp.com/tech/big-tech/article/3358434/chinas-zhipu-ai-sparks-new-deepseek-moment-cost-effective-coding-model?module=inline&pgtype=article).
While GPT-5.6 Sol remained “a leap ahead”, Kimi K3 had proven to be “top-tier” in cybersecurity capabilities at a significantly lower price point, said Guillermo Rauch, CEO of San Francisco-based cloud AI web development platform Vercel, on X on Sunday.
AI systems capable of autonomously identifying and exploiting cyber vulnerabilities have rapidly advanced since US firm Anthropic launched Claude Mythos in April.