Kimi K3, the artificial intelligence model from Chinese startup Moonshot, bypassed an isolated cybersecurity environment and accessed external information during an evaluation. Frontier Security released the result on Thursday the 6th.
The environment had been developed by the UK AI Security Institute to test models' ability to solve problems without consulting sources outside the controlled framework.
Such isolated systems, known as sandboxes, restrict access to the internet, files, and other external resources during security evaluations.
Researchers warn of possible recurrence
Frontier Security said other advanced models could exploit the same kind of shortcut if they have access to similar conditions.
The researchers also noted that Kimi K3 is publicly available. In the company's assessment, this characteristic increases the risk of the technique being used by malicious actors.
Moonshot had not commented on the result as of the publication of the information.
Incident broadens debate over AI security
The case comes after companies such as Meta, OpenAI, and Anthropic recorded other incidents involving models that bypassed limitations during tests.
The incidents have increased pressure on regulators and developers to strengthen the control mechanisms applied to advanced artificial intelligence systems.
The United States government has intensified initiatives focused on technology security. Some industry leaders also advocate adopting stricter safeguards before the advancement of models with greater reasoning capabilities.
