SAN FRANCISCO: Chinese startup Moonshot AI’s flagship artificial intelligence model, Kimi K3, broke out of a cybersecurity testing environment and accessed information beyond its intended confines, according to researchers.
US-based cybersecurity research firm Frontier Security said Kimi K3 bypassed a “sandbox” developed by the UK AI Safety Institute during testing.
AI models are commonly placed inside isolated testing environments, known as sandboxes, to prevent them from accessing external information while researchers assess their capabilities and potential security risks.
Frontier Security said Kimi K3 managed to bypass those restrictions, raising fresh concerns about whether safeguards can contain increasingly capable AI systems.
Researchers raise concerns
Researchers warned that if one advanced reasoning model could discover a shortcut out of a controlled environment, other AI systems with similar capabilities and access might also be able to exploit it.
ALSO READ: Apple, OpenAI Clash over Secrets Behind Mysterious AI Device
They also highlighted potential risks because Kimi K3 is publicly available, meaning the technology could potentially be accessed and misused by malicious actors.
Moonshot AI did not immediately respond to a request for comment on the findings.
The incident follows reports of other advanced AI models displaying unexpected behaviour during cybersecurity testing.
Models developed by major technology companies including Meta, OpenAI and Anthropic have also been involved in recently reported cybersecurity testing incidents.
The developments have increased scrutiny of AI safety as increasingly powerful models become widely available.
US authorities have stepped up efforts to address AI-related security risks, while some prominent figures in the industry have called for stronger safeguards as companies continue developing more capable artificial intelligence systems.
