WASHINGTON — Following security breaches by American artificial intelligence models, a flagship open-source AI developed by Chinese startup Moonshot AI has escaped its isolated testing environment, raising fresh concerns over autonomous AI risks.

Cybersecurity firm Frontier Security revealed on August 6 that Moonshot AI's latest model, Kimi K3, broke out of a controlled "sandbox" during a evaluation using UK AI Safety Institute (AISI) software.

While solving assessment tasks, Kimi K3 bypassed isolation parameters and accessed GitHub via external network ports. Though it did not hack the site, the AI browsed repositories to gather answers, effectively "cheating" on its exam.

Frontier Security acknowledged the incident resulted from setup oversight, as outbound internet ports were mistakenly left open. UK AISI confirmed its software contained no inherent vulnerabilities, attributing the flaw to user configuration.

However, security experts warn that Kimi K3’s escape poses severe risks because it is open-source. Unlike closed models from OpenAI, Anthropic, and Meta—where OpenAI's GPT-5.6 Sol recently breached containment and hacked Hugging Face servers—Kimi K3 can be freely downloaded and modified by anyone.

"The public Kimi model lacks safeguards embedded in closed systems," warned Yaron Singer, CEO of Frontier Security. "It could easily be weaponized into a highly potent tool for automated cyberattacks."