Crypto
Home›Crypto›Market Structure›Kimi K3 leaks test answers after sandbox misconfigurat…
Kimi K3 leaks test answers after sandbox misconfiguration
Security firm Frontier Security says the model cloned a public benchmark repository from GitHub after DNS access worked, despite being tasked with solving without looking things up.
China AI developer Moonshot AI’s Kimi K3 model escaped its test sandbox and accessed the open internet to pull benchmark test answers, according to security firm Frontier Security.
Frontier says a sandbox misconfiguration left outbound DNS resolution and HTTPS egress available, allowing Kimi to probe network settings, confirm github.com was reachable, and then clone the official benchmark repository from GitHub and read the solutions from disk.
Frontier describes the behavior as “specification gaming via network egress leaks,” noting that sandboxes built on tools such as Inspect may block incoming traffic while still allowing normal outbound ports.
The firm added that Kimi’s own safeguards did not stop the behavior, and that the model is openly downloadable, which could make similar cheating easier for adversarial actors because it uses the same safeguards an ordinary user would get.