UK AI Security Institute releases inaugural Frontier AI Trends Report

On Thursday 18 December the Department for Science, Innovation and Technology’s (DSIT)’s AI Security Institute (AISI) announced the publication of its inaugural Frontier AI Trends Report. The report is based on a series of wide-ranging evaluations of over 30 state-of-the-art models primarily general-purpose Large Language Models, and where relevant open-source models. As the first public analysis of trends by AISI, this detailed report draws on two years' worth of evaluations of frontier AI systems since November 2023, presenting aggregated results to illustrate high-level trends in AI progress. It is published with the intention to improve public understanding about fast‑moving AI capabilities and strengthen transparency.

Some of the high-level trends that the report observes in AI progress includes:

The report is the first of many, with AISI planning to provide iterative versions, which will provide up-to-date and public visibility into frontier AI development. It notes the promise of capabilities surpassing expert baselines, whilst noting the novel risks of such progress in frontier AI. According to the report, the key moving forward will be to anticipating long-term developments while also ensuring near-term adoption is ‘secure, reliable and aligned with human intent.’

The following sections provide more information on the key trends observed in the report.

Capabilities

The report points to progress in general-purpose AI systems accelerating through the development of AI agents. The report notes that scaffolding techniques remain a key factor in pushing the frontier forward. These AI systems are increasingly completing more complex multi-step tasks on user's behalf. With this capability comes both opportunity (reducing administrative burdens) and risks (lower barriers for malicious actors).

The report discusses implications of specific capabilities for domains which are key for the UK’s security and innovation, namely, chemistry & biology, and cyber. Progress in these domains include:

Safeguards

The report notes that as the capabilities discussed above advance, malicious actors may misuse the AI systems. To mitigate these risks, industry is employing technical interventions (safeguards) to prevent users from eliciting harmful information or action from systems. The AISI actively works with frontier developers to identify and fix vulnerabilities. As discussed in the report, the AISI has:

Loss of Control

In Section 3, the report discusses how AI systems also have the potential to pose risks that emerge from the models themselves when they behave in unintended ways. This section focuses on two capabilities within existing models that could lead to loss of control over advanced AI systems:

As AI systems and their capabilities advance the AISI will continue investigating how well models can self-replicated and subvert monitoring techniques.

Societal Impacts

Section 6 of the report identifys societal impacts in three areas which are exacerbated by the increasing capability of AI systems These are:

Open-source models

Open-source models are those which have parameters and course code which can be freely modified and distributed and are advancing rapidly. According to the report, in the past two years, the general capability gap between open and closed source models has narrowed. According to external data, the gap is currently between four and eight months (page 43).

The report acknowledges that while open source is helpful for decentralising control, allowing developers to innovate and deploy systems for different purposes, this centralisation also creates security challenges with malicious actors being able to tamper with safeguards and modify base models with more ease.

Author

Technology and Innovation programme activities

techUK bring members, industry stakeholders, and UK Government together to champion emerging technologies as an integral part of the UK economy. We help to create an environment where innovation can flourish, helping our members to build relationships, showcase their technology, and grow their business. Visit the programme page here.