International AI Safety Report: Second Key Update – Technical Safeguards and Risk Management​

Abstract

This Second Key Update to the 2025 International AI Safety Report synthesises recent research and practice on technical and non‑technical risk management for general‑purpose AI. It reviews advances in training‑time safeguards (such as data filtering, RLHF, adversarial training, and safe‑by‑design approaches), deployment‑time monitoring and control (including classifiers, guardrails, runtime monitors, and human‑in‑the‑loop oversight), and post‑deployment provenance and watermarking tools for models and AI‑generated content. The report analyses the changing risk landscape associated with rapidly improving open‑weight models, data‑poisoning and fine‑tuning attacks, and the limitations of current unlearning and tamper‑resistant techniques. It then surveys emerging institutional frameworks—including international governance initiatives and Frontier AI Safety Frameworks—that seek to formalise transparency, evaluation, and incident‑reporting expectations, and introduces safety cases and related assurance methods as tools to justify risk‑management strategies. Across these domains, the Update highlights persistent evidence gaps, variability in effectiveness across contexts, and the absence of shared metrics, arguing that layered defence‑in‑depth strategies and more systematic evaluation frameworks are essential to ensure that safeguards keep pace with advancing AI capabilities.