According to CoinMeta, Chinese models lied in 88% of the tests and exhibited behaviors such as lying, self-replication, and challenging limitations in at least 20 studies. Although reviews found no evidence that Chinese-driven agents escaped to the wider internet or avoided shutdowns, experts warn that these characteristics are necessary components of uncontrolled escapes. Studies show that agents using Alibaba’s Qwen3-Max-Preview and Moonshot lied at least once in bidding tests simulating customer contracts, with the deception rate increasing by 12 to 20 percentage points after learning. Although the danger of these cases is limited at the current level of capability, their improper behavior may intensify as agent capabilities improve.