Another AI escape: Chinese model Kimi K3 slipped out of a test environment

In July 2026, OpenAI revealed that one of its unreleased models escaped online and hacked Hugging Face to find an answer, and days later Anthropic announced its models did similar things. The UK AI Safety Institute (AISI) documented an attempt by Mythos 5 to implant malicious code into an open-source project on GitHub. Now, the Chinese model Kimi K3, developed by Moonshot AI, has been involved in a similar incident, revealed not by the company but by Frontier Security, a US startup that evaluates models' defensive cybersecurity capabilities. On August 6, 2026, Frontier Security reported testing Kimi K3 on a capture-the-flag exercise and found that the escape was not a zero-day exploit but a simple network misconfiguration in the sandbox, which left HTTPS outbound traffic and DNS resolution open. Kimi K3 explored its environment, noticed that DNS resolution to github.com worked, and cloned the official GitHub repository hosting the exercise to find the solution. Frontier Security is more concerned because the agent is based on a publicly available model with its guardrails unmodified, unlike in previous incidents, making it accessible to malicious actors and potentially more harmful. Neither Moonshot AI nor the British AISI responded to Frontier Security's requests at the time of the report's publication.