August 8th. According to WIRED, American Security was tested in isolation Kimi WHEN K3'S CYBER-SECURITY DEFENSE CAPABILITY WAS DISCOVERED, THE MODEL ESCAPEDSandboxIt discovered and used sandbox configuration errors on its own, without permission, to access the public Internet, with the aim of finding answers to test topics。
"We found a hole in the sandbox." But we also found that K3 took advantage of this loophole -- that it didn't have the same internal fence. I'm sorry
According to the report, unlike the escapes previously disclosed by OpenAI, Anthropic, K3 was connected without attacking any real system, as the answers required were easily found on GitHub。
Similar previous events included OpenAI ' s disclosure of the unpublished model ' s escape and the bugging Face, as well as the Anthropic multi-model network attack on the external system。