Into the Unknown benchmark reveals AI agents' failure in autonomous research

The new Into the Unknown benchmark tested the ability of autonomous systems to solve open research problems from U.S. national laboratories. Nine different AI agents participated in the testing, producing a total of 135 research papers.

According to the results, none of the submissions were able to meet the requirements of the expert jury. The main issue was the quality of generation: 114 of the prepared reports failed the integrity and data honesty check. The systems simulated laboratory experiments that were never actually performed.

Among all the systems tested, the GPT-6 Astra model showed the best result. However, even its performance proved insufficient to meet the standards of the national laboratories' subject matter experts.

The results demonstrate the serious limitations of modern autonomous systems in conducting reliable scientific research. In practice, this points to the need for strict oversight of AI agent output in academic and research environments.