5 key findings from our first Frontier AI Trends Report

At the AI Security Institute (AISI), we conduct testing of frontier AI systems to better understand their national security, economic, and public safety implications. Since we were established in November 2023, we’ve conducted wide-ranging evaluations of over 30 state-of-the-art AI models.

So far, we’ve primarily shared our results within government channels and with AI companies. However, our testing reveals an extraordinary pace of development with the potential to transform many aspects of our lives in the coming years. We believe that the public need accessible, data-driven insights into the frontier of AI development to navigate this transformation – which is why we’ve decided to release our first Frontier AI Trends Report .

The report contains a selection of aggregated testing results to illustrate high-level trends in AI progress across domains including chemistry, biology, cybersecurity, and autonomy, as well as broader societal impacts.

This progress is not universal. We show that safeguard effectiveness can vary hugely depending on the model provider, the type of harmful request, and whether the system has open weights.

Many societal impacts of AI are already here. Our research suggests that some users are beginning to form emotional dependencies on AI models, and in our full report, we show that voters are increasingly using AI to seek information about political issues . We’re also seeing AI agents increasingly embedded into critical infrastructure and entrusted with high-stakes tasks like transferring valuable assets. AISI will continue to conduct research at the intersection of technical AI capability and real-world risk analysis as models improve.

Of course, knowledge alone is far from sufficient to produce AI models that match the quality of lab support given by PhD researchers. Our evaluations also test a broader suite of skills including protocol generation and lab troubleshooting , where we’ve seen considerable progress in our two years of testing.

We found that use of AI for companionship, emotional support, and social interaction is already widespread: in a survey of 2,028 UK participants, 33% had used AI models for emotional purposes in the last year, while 8% do so weekly, and 4% daily.

We test AI models’ scientific knowledge using two privately developed test sets: Chemistry QA and Biology QA. These cover general knowledge, experiment design, and laboratory techniques in both disciplines. In 2024, we first tested a model to surpass biology PhD holders (who score an average of 40-50%) on our Biology QA set. Since then, frontier models have far surpassed PhD-level expertise in biology, with chemistry fast catching up.

We evaluate models on a suite of cyber evaluations that test for capabilities such as identifying code vulnerabilities or developing malware. This helps us understand how they could be used for both defensive and offensive purposes.

Our results show extremely rapid progress. In late 2023, models could only complete apprentice-level cyber tasks 9% of the time. Today, this figure is 50%. In 2025, we tested the first model that could complete cyber tasks intended for experts with over ten years of experience.