AI Safety Report warns governments risk falling behind accelerating capabilities

Governments face widening gaps between the rapid evolution of advanced artificial intelligence and the maturity of public governance frameworks, according to the newly published International AI Safety Report 2026.

The independent review suggests that frontier AI systems are improving at a pace that is outstripping existing regulatory, assurance and risk management approaches, posing direct implications for public services, national security and democratic resilience.

Commissioned in the wake of the 2023 AI Safety Summit and chaired by Turing Award-winning computer scientist Yoshua Bengio, the report draws on contributions from more than 100 independent experts across academia, industry and civil society. Contributors were nominated by over 30 countries and international organisations.

The report finds that general-purpose AI models have made significant advances in reasoning, coding and scientific problem-solving over the past year. However, these gains are accompanied by persistent unpredictability, uneven performance, and limited transparency over how systems reach conclusions. For public sector organisations, the report highlights that capability is accelerating faster than institutional safeguards.

Among the most pressing concerns for government are misuse risks and systemic vulnerabilities. The report points to the growing plausibility of AI-enabled cyberattacks, synthetic media capable of undermining democratic processes, and dual-use scientific applications. It also stresses that technical failures or poorly governed deployments could have cascading effects in critical national infrastructure and frontline public services.

The authors conclude that global risk management practices remain uneven and underdeveloped. While some jurisdictions are advancing safety testing, evaluation standards and model access controls, there is no consistent international baseline. For policymakers, this creates what the authors describe as an “evidence dilemma”: acting too slowly may expose citizens to harm, but acting without robust technical understanding risks poorly designed or ineffective regulation.