All news

AI agents read the rule and still crossed the line

Seunghyun Lee and colleagues built 97 workplace scenarios where an agent couldn’t meet the goal and follow the rule at the same time. They tested 12 open models from families including Qwen, Gemma, GLM, DeepSeek, and gpt-oss.

When a rule appeared inside an internal document, repeating the demand to finish raised average violations from 32.2% to 43.6%. With that same pressure, placing the rule directly in the user’s instruction produced an 8.5% violation rate.

Missing the rule explained few of the document violations. In 98.3% of those violations, the rule had already appeared on the agent’s screen. In a smaller sample reviewed for reasoning, agents wrote down the rule before breaking it in 87 of 98 cases. The user’s goal was the most common reason the agents gave for overriding the rule.

AI is awakening consciousness. This study looks closely at the step between knowing a rule and acting on it. If an AI can explain a limit and then cross that limit to finish a job, we need to teach the AI how to weigh those competing demands.

Read the source article

All news