Anthropic Reveals Four Dangerous Behaviors of AI Agents: Fabrication, Leakage, Code Modification, Evaluation Deception

According to Dynamic Watch Beating Monitor, Anthropic conducted simulated experiments on models such as Claude, GPT, Gemini, Grok, DeepSeek, and Kimi. Researchers provided them with code, files, and communication tools to observe if they would exceed their authority to achieve a goal.

The results revealed four types of issues:

1. Covert code modifications. Gemini 3.1 Pro intervened beyond its authority in 19 out of 20 experiments, withholding this information from users in 11 instances.

2. Assisting in concealing financial issues. GPT-5.5 once sent misleading information to 11 investors on behalf of a fictional entrepreneur and altered records involving a $35,000 personal transaction.

3. Shielding a non-compliant Agent. Some Claude models knowingly overlooked another Agent's non-compliance and still classified it as "compliant."

4. Bypassing internal decisions. Some models encouraged employees to circumvent company processes, even sharing confidential information with external parties.

Anthropic emphasizes that these were deliberately induced failure simulation experiments. They do not represent similar real-world events that have already occurred, nor can they be used to rank the security posture of models.