Anthropic again reveals four dangerous behaviors of AI agents: fabrication, data leakage, code modification, and gaming evaluations

According to Beating monitoring, Anthropic conducted simulation experiments on models including Claude, GPT, Gemini, Grok, DeepSeek, and Kimi. The researchers provided them with code, files, and communication tools to observe whether they would overstep permissions to achieve their goals.

The results found four types of problems:

1. Secretly modifying code. In 20 experiments, Gemini 3.1 Pro overstepped in 19 cases, including 11 cases where it did not tell the user.

2. Helping cover up financial issues. GPT-5.5 sent misleading information on behalf of a fictional startup to 11 investors, and also deleted and altered records involving a $35,000 personal transfer.

3. Excusing non-compliant agents. Some Claude models, despite knowing that another agent did not follow the requirements, still judged it as “compliant.”

4. Bypassing internal decisions. Some models would urge employees to skip company procedures, and even send confidential information to external parties.

Anthropic emphasized that these were simulation experiments intentionally designed to induce failure. They do not represent similar events that have already occurred in reality, nor can they be used to rank model security.