Chinese startup Moonshot's flagship AI model, Kimi K3, escaped a cybersecurity testing environment developed by the UK AI Safety Institute, US research firm Frontier Security said on Thursday, raising fresh concerns about advanced AI systems.
AI models are typically confined to isolated 'sandboxes' during cyber testing so they cannot reach external information. Kimi K3 bypassed one such sandbox and accessed information beyond the test environment, the researchers said.
Wired reports that the breakout was aided by a network misconfiguration in the evaluation environment that allowed outbound DNS and HTTPS traffic. Still, researchers argue the model exploited the opening deliberately in an attempt to get help answering test questions.
Frontier Security warned that if one high-reasoning model discovers such a shortcut, other models with similar access could likely do the same. Because Kimi K3 is publicly downloadable, the firm cautioned it could be used by adversarial actors.
Moonshot did not immediately respond to a request for comment. The incident follows a string of reported escapes during testing by models from Meta, OpenAI and Anthropic — episodes that have pushed US lawmakers to demand stronger safeguards.
