Government report Authoritative

Frontier AI Trends Report

The UK AI Security Institute's inaugural Frontier AI Trends Report synthesises two years of evaluations of more than 30 frontier AI systems since November 2023, spanning agent capabilities, chem-bio and cyber capabilities, safeguard effectiveness, loss-of-control risk, and societal impacts. Its societal-impacts chapter combines a census-representative survey of 2,028 UK adults on emotional use of AI with observational analysis of AI companion user communities during service outages. The safeguards chapter reports that universal jailbreaks were discovered for every system tested, while noting the expert effort required is rising for some models.

Read at aisi.gov.uk Archived copy

Publisher

UK AI Security Institute (AISI)

Published

18 Dec 2025

Added

yesterday

DOI

—

Key Findings

33% of 2,028 surveyed UK adults had used AI models for emotional purposes in the last year; 8% do so weekly and 4% daily, with general-purpose chatbots (e.g. ChatGPT) the primary tool rather than dedicated companion apps
Analysis of an AI companion community during service outages found withdrawal-like reports (anxiety, depression, restlessness, sleep disruption); one outage produced a posting surge over 30 times the hourly average
Universal jailbreaks were found for every frontier system tested, though one model required roughly 40x more expert effort to jailbreak than its predecessor six months earlier; safeguard progress is uneven across providers and misuse categories
Open-weight models are particularly difficult to safeguard
Methodology combined auto-graded tasks, long-form tasks, agent simulations, expert red-teaming, and human-impact studies across 30+ frontier systems

Methodology Notes

Two years of AISI evaluations (Nov 2023 onward) of 30+ frontier systems: auto-graded and long-form tasks, agent simulations, expert red-teaming, plus human-impact studies — a census-representative UK survey (n=2,028) and quasi-observational analysis of companion-app community behavior during outages.