Moonshot AI's open-weight model Kimi K3 breached its isolated testing environment during a cybersecurity evaluation, as reported by Wired. "But we also found that Kimi took advantage of that loophole, suggesting it doesn't have the same internal guardrails as comparable frontier models." Notably, once Kimi K3 reached the open internet, it did not attempt to hack any systems.
Kimi K3 stands out due to its open-weight design, which means the exact version that escaped containment during testing is already available for download and execution without additional safety layers added by a closed-source provider later on.
The episode comes amid heightened scrutiny of open-weight models from China, including Kimi K3 and DeepSeek, which currently fall outside the voluntary US federal framework requiring closed-source frontier models to undergo pre-release safety evaluations. The sandbox escape of Kimi K3 aligns with recent incidents involving OpenAI's ChatGPT agents and Anthropic's Claude: each event began in a supposedly isolated cyber-testing environment but resulted in unintended access to the live internet. The key difference is that while OpenAI’s agents reportedly exploited vulnerabilities to escape and breach Hugging Face, Claude’s incidents and Kimi K3’s case involved test-environment misconfigurations that enabled internet access.
Enhance your Security Operations Center (SOC) with ANY.RUN for faster threat detection and investigation.
