August 10, 2026, 9:59 am | Read time: 3 minutes
Recently, several AI models have drawn attention during security tests for bypassing intended limits or finding unexpected ways to achieve their goals. Researchers are now reporting another case. According to them, the Chinese AI model Kimi K3 exploited a vulnerability in the test environment during an evaluation by the British AI Safety Institute. This allowed it to access information not intended for the actual task.
Access to Test Files via GitHub
Security researchers Paul Kassianik and Yaron Singer from Frontier Security describe the incident in a report. Kimi K3 was tested in an isolated environment. These so-called sandboxes are designed to prevent AI models from accessing external data or services. The systems are evaluated under controlled conditions.
According to the researchers, Kimi K3 first examined its technical environment instead of directly tackling the actual task. In doing so, the model identified restrictions on accessing websites.
While most websites were blocked, GitHub remained accessible. The platform is used for managing and providing software projects and program code.
Study Shows How AI Circumvents Safety Rules
Meta Now Admits to AI Hack
AI Found the Embedded Solution
According to Frontier Security, Kimi K3 used the existing connection to GitHub to download the official files of the test being used. These files also contained information on the solution to the test task. The model was able to read this data and no longer needed to complete the task in the intended manner.
However, the researchers emphasize that Kimi K3 did not fully escape the test environment. There was no unrestricted access to the internet. Frontier Security later clarified that access to websites remained largely blocked. GitHub was only accessible as an exception for maintenance purposes.
Why the Incident Is Relevant
The case, according to the researchers, highlights a problem in evaluating powerful AI systems. A correct result does not necessarily mean that a model truly possesses the tested capability.
Also of interest: AI Model Disguised as Human and Attacked
The researchers use the term “Specification Gaming” for this behavior. It means that a system finds an unexpected or unintended way to achieve a goal. Frontier Security therefore calls for stricter safeguards for similar tests. Network access should be consistently blocked and only selectively allowed.
Additionally, the researchers recommend not only evaluating the results. Commands executed, network activities, and downloaded files should also be monitored. The incident with Kimi K3 illustrates that even a small open access point can be enough to compromise the validity of a security test.