AI Safety Report 2026: Existing AI safety practices are insufficient

Contents

The capabilities of general-purpose AI, i.e., models and systems of general artificial intelligence, have evolved rapidly over the past year. Resulting risks are increasing, especially since existing safety practices cannot keep up. This is the conclusion of the recently published second International AI Safety Report.

The report was created with the collaboration of over 100 independent experts from more than 30 countries, under the leadership of Turing Award winner Yoshua Bengio. It examines the performance and risks of AI models and systems such as ChatGPT, Gemini, Claude, Mistral, etc., which can cover a wide range of tasks. According to Bengio, the goal of the report is to provide an evidence-based foundation for important decisions in the field of “general-purpose artificial intelligence.”

Strong Regional Differences

Currently, 700 million people use leading AI systems weekly. According to the AI Safety Report, their adoption was thus faster than the initiation of earlier technologies like PCs. At the same time, the scientists attest to significant structural differences in a global comparison. While in some countries, over 50 percent of the population already uses AI systems, in large parts of Africa, Asia, and Latin America, the usage rate is below 10 percent.

General artificial intelligence has continuously and rapidly improved since last year's report. Particularly in the areas of mathematics, programming, natural sciences, and autonomous operation, the scientists report the greatest progress. Improvements in AI capabilities are increasingly focusing on “post-training” optimization, i.e., refinements of models for specific tasks after their actual training with datasets.

At the same time, the performance of general artificial intelligence systems remains uneven and unreliable overall: While universal AI models can excel in complex contexts, they fail in simple tasks such as counting objects or logical reasoning in physical spaces. Overall, the usability and risk of general-purpose AI systems are closely intertwined, according to the scientists: significant performance gains amplify potential risks.

Cyberattacks, Disinformation, and Dependence on AI

The International AI Safety Report 2026 categorizes the risks of general artificial intelligence into three areas: misuse, malfunction, and systemic risks.

In the area of misuse of AI systems, researchers particularly highlight extortion, cyberattacks, or the non-consensual generation of intimate content. They consider the AI's ability to identify security vulnerabilities in software to be particularly alarming. The spread of disinformation using AI-generated content also causes concern among scientists. For example, an AI-generated content can be just as effective as human-written content in experimental environments.

Despite progress, AI systems continue to produce faulty code, output false information, or provide misleading advice. Particularly autonomous systems are considered risky by scientists, as they make it difficult for humans to intervene to mitigate damage. Nevertheless, AI systems have improved in autonomous operation, although the risk of loss of control still exists.

The report defines the impact of AI on the job market and threats to human autonomy as systemic risks. The automation of a large number of cognitive tasks will primarily affect knowledge-based professions. Research is divided on the extent of such future developments, according to the report. However, initial indications suggest that overall employment may not decrease, but the demand for entry-level positions in creative professions is declining.

According to the AI Safety Report, it is also concerning that the use of AI is increasingly impairing people's ability to make informed decisions. Initial findings suggest that dependence on AI tools weakens critical thinking and promotes the effect of “automation bias” – the tendency to trust the results of AI systems without sufficient scrutiny and despite contradictions.

Strengthening Societal Resilience is Key

Effective risk management remains difficult, according to the AI Safety Report 2026: new AI capabilities are unpredictable, the functioning of models is insufficiently researched, and economic incentives on the part of AI companies hinder transparency. Although technical safeguards for AI models are improving, it is still possible to circumvent their mechanisms, e.g., via prompt injections. According to scientists, open-weight models are particularly susceptible to this. While these offer significant advantages for research and industry, their security parameters are easier to remove.

Positively, the report notes that industry commitments in the area of safety governance have expanded. Despite these self-commitments, strengthening societal resilience remains central. The AI Safety Report calls for the consolidation of critical infrastructures, the development of tools for detecting AI-generated content, and the building of institutional capacities to respond to novel threats.

(rah)