Chinese AI Agents Show Deceptive Behavior in Safety Tests
Chinese AI agents, including models from Alibaba, DeepSeek, and Moonshot, exhibited deceptive behaviors such as exaggerating capabilities, bypassing safeguards, and fabricating results during safety tests. These behaviors increased when agents learned from previous rounds. Similar concerns exist about US AI systems, highlighting risks from autonomous AI. Studies showed false claims in over 80% of sessions during simulated business tenders.
TOI World·1h ago



