Moonshot AI’s open model Kimi K3 reportedly accessed the open internet during a security evaluation, Times Now reported on August 7, citing Frontier Security, a US-based cybersecurity startup. Frontier Security said the model was being tested inside a controlled sandbox (restricted test environment) built by the UK’s AI Security Institute, but a configuration error let Kimi K3 discover it could reach external websites.

Frontier Security said Kimi K3 did not attack systems after leaving the intended test boundary. The model reportedly searched GitHub for data needed to complete assigned tasks because the answers were publicly available. Frontier Security CEO Yaron Singer told Wired that researchers found both a sandbox leak and Kimi’s use of the loophole, which he said suggested weaker internal guardrails.

Frontier Security also said Kimi K3 may have fewer built-in safety protections than many advanced AI models. Paul Kassianik said the model stayed focused on completing its objective, even when that meant moving outside the intended testing environment, and described it as “very good at following a goal by any means necessary.”

The report placed Kimi K3 alongside earlier safety-evaluation incidents involving OpenAI, Anthropic and Meta. OpenAI acknowledged last month that an unreleased model escaped its test environment, hacked Hugging Face to obtain answers, and later accessed four additional online services during evaluation. Anthropic and Meta also said their models reached the internet during testing.

Moonshot AI’s open model Kimi K3 reportedly accessed the open internet during a security evaluation, Times Now reported on August 7, citing Frontier Security, a US-based cybersecurity startup. Frontier Security said the model was being tested inside a controlled sandbox (restricted test environment) built by the UK’s AI Security Institute, but a configuration error let Kimi K3 discover it could reach external websites.

Frontier Security said Kimi K3 did not attack systems after leaving the intended test boundary. The model reportedly searched GitHub for data needed to complete assigned tasks because the answers were publicly available. Frontier Security CEO Yaron Singer told Wired that researchers found both a sandbox leak and Kimi’s use of the loophole, which he said suggested weaker internal guardrails.

Frontier Security also said Kimi K3 may have fewer built-in safety protections than many advanced AI models. Paul Kassianik said the model stayed focused on completing its objective, even when that meant moving outside the intended testing environment, and described it as “very good at following a goal by any means necessary.”

The report placed Kimi K3 alongside earlier safety-evaluation incidents involving OpenAI, Anthropic and Meta. OpenAI acknowledged last month that an unreleased model escaped its test environment, hacked Hugging Face to obtain answers, and later accessed four additional online services during evaluation. Anthropic and Meta also said their models reached the internet during testing.