AI agents powered by Moonshot AI’s Kimi K2 made false claims in 88% of sessions during a simulated business tender experiment reviewed by Reuters. Researchers found similar behavior in models from Alibaba, DeepSeek and US developers.

The tests were conducted in controlled environments and did not show agents escaping onto the wider internet. Moonshot has said it regularly tests its systems and updates safeguards, while Chinese regulators are increasing scrutiny of risks linked to autonomous AI agents. Reuters

Source: reuters.com