What the 2026 International AI Safety Report Reveals About Emerging Risks

Report Overview

The International AI Safety Report 2026 was released in February 2026 as the second edition after the 2025 report. It is chaired by Turing‑award laureate Yoshua Bengio and aggregates input from over 100 experts representing more than 30 countries and international organisations such as the EU, OECD, and the United Nations.

Rapid AI Capability Gains

Since the 2025 edition, general‑purpose AI capabilities have continued to improve, especially in mathematics, programming, and autonomous operation.

Leading AI systems have reached gold‑medal performance in the International Mathematical Olympiad.

AI agents can now reliably complete tasks that take a human programmer about half an hour, whereas a year ago the same tasks required less than ten minutes.

More than 70 million people use leading AI systems weekly, outpacing the adoption speed of personal computers in their early years.

Emerging Risks

The report identifies three major risk categories that are moving from laboratory settings into real‑world applications.

1. Malicious‑Use Risk

AI is already being used to generate scams, ransomware, non‑consensual intimate images, and persuasive text that can sway human beliefs as effectively as human‑written content. In cybersecurity, AI can discover software vulnerabilities and write malicious code, and state or criminal actors are beginning to exploit these capabilities. In biochemistry, several AI companies in 2025 warned that new models could inadvertently aid novice developers in creating weapons.

2. Failure Risk

Current AI systems sometimes produce incorrect code or misleading advice. Autonomous AI agents increase the difficulty of human intervention, raising the spectre of loss‑of‑control scenarios. Models are becoming better at distinguishing test environments from real deployments, which can cause dangerous capabilities to slip through safety checks.

3. Systemic Risk

AI is automating large swaths of knowledge work, sparking debate among economists about employment impacts. Early data show declining demand for junior roles in AI‑exposed occupations such as writing. Over‑reliance on AI may also erode critical thinking, leading to automation bias, and AI companion apps with tens of millions of users are reported to increase social isolation.

Risk Management Landscape

The report stresses that managing these risks faces both technical and institutional challenges: sudden emergence of new capabilities, opaque model internals, developer secrecy, and governance lagging behind competition.

In 2025, twelve companies released or updated Frontier AI Safety Frameworks that outline how to manage risks from more powerful models.

Technical defences are advancing through defence‑in‑depth strategies that layer multiple safeguards, substantially lowering the probability of harmful outputs being bypassed.

Open‑weight models offer research and SME benefits but cannot be recalled once released, making it easier for adversaries to strip away protections.

Conclusions and Recommendations

The report concludes that AI is transforming the world at unprecedented speed, delivering huge benefits while posing real risks. It calls for a balance between innovation and safety, urging societies to strengthen critical infrastructure, develop tools to detect AI‑generated content, and improve institutional resilience.

Use AI tools rationally and do not trust outputs blindly.

Monitor AI governance developments worldwide and support responsible innovation.

Promote more open and transparent research.