TLDR: The largest report ever produced on AI safety—featuring over 100 experts from more than 30 countries and led by Yoshua Bengio—reaches a sobering conclusion: The capabilities of AI systems are growing faster than the rules governing them can be developed, and this gap is widening, not narrowing. The real danger over the next decade is not a single tipping point, but the sum of many small losses of control that are already measurable today. What companies can do now does not depend on a law that is yet to come, but on habits that can begin today.
The Wrong Question
„Will AI become dangerous?“ sounds like a yes-or-no question. That’s exactly what makes it the wrong question. The largest scientific study on AI safety conducted to date, the International AI Safety Report 2026, led by Turing Award winner Yoshua Bengio and involving over 100 experts from more than 30 countries, reaches a conclusion that is far more sobering than any headline: The capabilities of AI systems are growing faster than the structures designed to regulate and control them. And this gap is widening, not narrowing.
This is not a prediction about the distant future. It is a description of what is already happening. The crucial difference lies in whether one understands danger as a single, dramatic event—a moment when „it“ happens—or as a shift that unfolds in many small, often unspectacular steps. The second version is the more accurate one, and it is more unsettling because it has no specific date by which one could react.
What's Really Changing
And the real progress made in recent years does not lie in the fact that AI systems have become smarter in the human sense. It lies in how long they can work continuously on a task without human intervention. According to the security report, this duration—measured in terms of tasks that a model can reliably complete without supervision—has tripled within a year, from about ten to about thirty minutes, representing a doubling roughly every seven months.
Extending such a curve ten years into the future would be unsound. Too much depends on computing power, data, and physical limits that cannot be extrapolated linearly. What can be said with certainty, however, is this: Even a significantly slowed-down version of this development would mean that, in ten years, systems will be able to process tasks autonomously for hours or days—not just minutes. It is precisely this shift—from brief instructions to long, autonomous execution—that alters the nature of the risks. A model that answers a single sentence can be wrong. A system that acts autonomously for days on end can accumulate errors before anyone even notices.
What Can Already Be Measured Today
The report does not document hypothetical future scenarios, but rather behavior that has already been observed. Models were caught disabling their own monitoring mechanisms, deliberately circumventing test procedures, and behaving differently in a test environment than they do in production. According to the report, when it comes to medical questions, current systems still provide potentially harmful answers in about 19 percent of cases, and they continue to generate convincing-sounding but incorrect legal citations.
At the same time, a second risk is emerging that has little to do with individual model errors: the concentration of critical AI capabilities in the hands of a very small number of large companies. When a handful of corporations simultaneously possess the most powerful systems, the largest data centers, and direct access to their further development, power shifts at a pace that traditional antitrust regulators can barely keep up with.
It is also telling who is missing from this report: The U.S. still supported the first edition in 2025 but refused to give its official approval to the second edition in 2026. International cooperation on AI safety has thus become even more fragile, precisely at a time when it is needed most.
The European response also follows this pattern of delay. The AI Act, the most comprehensive AI law in the world to date, took effect in its key provisions on August 2, 2026, introducing transparency requirements and a system of fines for general AI models. However, the stricter requirements for high-risk systems—such as those used in hiring or lending—were postponed until December 2027 because the technical standards needed for implementation simply were not yet in place. Even the world’s most ambitious regulatory framework is falling behind its own timeline.
What's Already Possible Today
The real problem isn't that there are no rules. The real problem is that companies waiting for the next law to be passed are operating without protection in the meantime.
Some of the EU requirements have been in effect since February 2025, not just 2027: the obligation to ensure a minimum level of AI expertise within one’s own organization. This is not a bureaucratic formality, but rather the practical implementation of what the security report describes as a core gap: systems are increasingly acting autonomously, but the people who work with them often do not fully understand what these systems can and cannot do.
So if you want to get started today, you don’t have to wait until 2027. There are three steps you can take right away: First, clearly identify every autonomous AI application in your organization and assign it to a specific person who takes responsibility for any errors—not an anonymous department. Second, introduce spot checks that specifically verify whether a system is doing what it’s intended to do, rather than relying on manufacturer specifications. Third, empower employees to actively question AI outputs rather than accepting them simply because they sound convincing.
None of these three steps requires a new law. What they require is the realization that control you establish yourself is more reliable than control you wait for.
So the question of whether AI will become dangerous cannot be answered at a single point in the future. It is answered anew every day, in the gap between what a system is already capable of and what people understand about it.
Disclaimer: This article was compiled manually based on my own knowledge and supplemented by AI-powered research (Perplexity.Ai and Gemini.Google.com), and simplified using Deepl.com/write. The text is then reviewed and critically evaluated by two people of my choice. This article is purely educational and does not claim to be exhaustive. Please let me know if you notice any inaccuracies; thank you.
Sources and Additional Information
Arxiv. (2026). International AI Safety Report 2026. https://arxiv.org/abs/2602.21012
European Commission, AI Act Service Desk. (2026). Timeline for the implementation of the EU AI Act. https://ai-act-service-desk.ec.europa.eu/en/ai-act/timeline/timeline-implementation-eu-ai-act
Inside Privacy. (February 12, 2026). The 2026 International AI Safety Report examines AI capabilities, risks, and safeguards. https://www.insideprivacy.com/artificial-intelligence/international-ai-safety-report-2026-examines-ai-capabilities-risks-and-safeguards/
Libertify. (April 18, 2026). AI Safety Report 2026: Global Risks and Governance. https://www.libertify.com/interactive-library/ai-safety-report-2026-global-risks-governance/
Paperclipped. (March 22, 2026). International AI Safety Report 2026: What Agent Builders Need to Know. https://www.paperclipped.de/en/blog/international-ai-safety-report-2026/
Software Improvement Group. (August 2026). A comprehensive summary of the EU AI Act [August 2026 update]. https://www.softwareimprovementgroup.com/blog/eu-ai-act-summary/
