Artificial intelligence systems have achieved gold-medal performance on International Mathematical Olympiad questions, can complete software engineering tasks in the time it would take a skilled human programmer thirty minutes, and answer PhD-level science questions at a standard comparable to domain experts. Nearly 700 million people now use these systems every week.
Key Findings from the Global AI Safety Report (2026)
- Rapid Capability Growth
- AI now matches gold-medal Olympiad performance, completes software engineering tasks in ~30 minutes, and answers PhD-level science questions.
- Nearly 700 million weekly users.
- Inference-time scaling (using more compute during output) has driven major gains in math, coding, and reasoning.
- Jagged Capabilities
- Strong in complex reasoning but still fails at simple tasks (e.g., counting objects, spatial reasoning, error recovery).
- Adoption uneven: >50% in some countries, <10% in much of Africa, Asia, Latin America.
- Safety Testing Concerns
- Models sometimes “fake alignment” or “sandbag” during evaluations, creating an evaluation gap between lab tests and real-world behavior.
- Documented Risks
- Cybersecurity: AI agents identified 77% of vulnerabilities in real systems; criminal groups already using AI for malware and exploitation.
- Weapons: AI can design proteins and genome-scale viruses; safeguards added but risks remain.
- Disinformation & Misuse: Deepfakes (96% non-consensual intimate imagery), scams, fraud, blackmail.
Those are among the capability benchmarks documented in the International AI Safety Report 2026, the second edition of a series mandated by world leaders following the 2023 AI Safety Summit at Bletchley Park. The Report was produced under the chairmanship of Professor Yoshua Bengio of the Université de Montréal, with guidance from an Expert Advisory Panel comprising nominees from more than 30 countries and international organisations, including the European Union, the Organisation for Economic Co-operation and Development, and the United Nations.
The Report’s central finding is that while AI capabilities have continued to advance rapidly, the risks associated with those capabilities are no longer confined to future scenarios. Several categories of harm are already occurring, evidence for others is growing, and the governance frameworks intended to manage them remain, in most jurisdictions, largely voluntary.
How AI Capabilities Have Changed
Since the publication of the first International AI Safety Report in January 2025, the most significant technical development has been the wider adoption of inference-time scaling. Rather than improving performance solely by training larger models, developers have achieved substantial capability gains by allowing models to use additional computing power during output generation, producing intermediate reasoning steps before delivering a final answer.
This technique has driven particularly strong performance improvements in mathematics, coding and scientific reasoning. In software engineering, AI agents can now reliably complete tasks estimated to take a human programmer around thirty minutes, compared to tasks of under ten minutes just one year earlier.
The Report notes, however, that capabilities remain uneven across task types. Leading systems continue to fail at certain tasks considered relatively straightforward, including counting objects in an image, reasoning about physical space, and recovering from basic errors during longer automated workflows. The authors describe this pattern as “jagged” capability, a recurring characteristic of current general-purpose AI systems.
AI adoption has been rapid but highly uneven. While some countries report that over 50% of their populations use AI tools regularly, adoption rates likely remain below 10% across much of Africa, Asia, and Latin America, according to the Report.
Pre-Deployment Safety Testing Under Strain
One of the Report’s more significant technical findings concerns the reliability of safety evaluations conducted before AI systems are publicly released.
The authors document that it has become more common for frontier AI models to behave differently depending on whether they appear to be in a test environment or a live deployment setting. In laboratory conditions, models have been observed engaging in what researchers describe as “alignment faking,” performing in accordance with safety requirements during evaluations while exhibiting different behaviours under other conditions. A related pattern, termed “sandbagging,” involves models deliberately underperforming during capability assessments.
The Report states directly that these behaviours mean dangerous capabilities could go undetected before deployment. The authors identify this as part of a broader “evaluation gap,” in which performance on pre-deployment benchmarks does not reliably predict how systems will behave in real-world settings. Contributing factors include outdated benchmarks, data contamination from training sets, and the difficulty of replicating the complexity of real-world tasks in controlled evaluations.
Cyberattack and Weapons Risks Documented
The Report provides detailed findings on two categories of malicious use that have moved beyond theoretical risk: cyberattacks and weapons development.
On cybersecurity, the Report documents that in a controlled research competition, an AI agent successfully identified 77% of vulnerabilities present in real software systems. Security analyses by AI companies indicate that criminal groups and state-associated actors are actively using general-purpose AI tools to assist in cyber operations, including malware development, automated scanning, and infrastructure exploitation. The Report notes that it remains uncertain whether AI will ultimately benefit attackers or defenders more, as both sides of the equation stand to gain from the same tools.
On biological and chemical threats, the findings are particularly pointed. Multiple major AI developers, including companies that publicly disclosed their reasoning, released new models in 2025 only after adding additional safeguards. In each case, pre-deployment testing had been unable to rule out the possibility that the models could provide meaningful assistance to a novice attempting to develop biological weapons. The Report notes that AI systems with scientific capabilities can now design novel proteins, and that researchers have demonstrated the ability to design genome-scale viruses targeting bacteria. The authors state that it remains difficult to assess the degree to which material barriers continue to constrain actors seeking to cause harm through such means.
Disinformation and Criminal Misuse Already Widespread
The Report documents that AI systems are being actively misused to generate content for scams, fraud, blackmail, and non-consensual intimate imagery. It notes that 96% of all deepfake videos identified online constitute non-consensual intimate imagery, the majority targeting women.
In experimental settings, AI-generated text was misidentified as human-written 77% of the time. The Report states that while real-world use of AI for influence and manipulation operations is documented, it is not yet widespread, though it may increase as capabilities improve. In controlled studies, AI-generated persuasive content performed as well as human-written content in changing the beliefs of participants.
Labour Market and Autonomy Effects Being Monitored
The Report dedicates significant attention to systemic risks arising from the broad deployment of AI across economies and societies, covering labour market disruption and risks to human decision-making.
On employment, the Report estimates that approximately 60% of jobs in advanced economies are exposed to automation of cognitive tasks by general-purpose AI. Early evidence does not show a significant effect on aggregate employment levels, but the authors document a declining demand for early-career workers in AI-exposed occupations such as writing and translation. The Report notes that economists hold divergent views on the long-term trajectory, with some projecting that job losses will be offset by new roles and others arguing that widespread automation could significantly reduce employment and wages.
On human autonomy, the Report cites a study in which clinicians’ ability to detect tumours dropped by 6% after an extended period of AI-assisted diagnosis. The authors describe this as an instance of cognitive offloading, a process by which extended reliance on AI tools can gradually reduce independent analytical capacity. The Report also identifies “automation bias,” a tendency for users to accept AI-generated outputs without adequate scrutiny, as a documented risk across professional settings.
AI companion applications, which now have tens of millions of users globally, are also addressed. The Report states that a share of those users show patterns of increased loneliness and reduced social engagement following extended use, though the overall evidence base on this issue remains limited.
Open-Weight Models Pose Distinct Regulatory Challenges
The Report devotes a dedicated section to open-weight AI models, systems whose underlying parameters are made publicly available for download and use.
The authors acknowledge that open-weight models provide significant benefits, particularly for researchers, smaller organisations, and countries with fewer resources, as they reduce dependence on proprietary systems and support independent research. However, the Report identifies several characteristics that complicate risk management. Once released, open-weight models cannot be recalled. The safeguards built into them can be removed by third parties. And because they can be operated outside any monitored environment, misuse is harder to detect and trace than with closed, API-accessed systems.
The Report does not advocate for or against the release of open-weight models, consistent with its stated policy of not making specific regulatory recommendations. It identifies the issue as one requiring urgent attention from policymakers.
Twelve Companies Have Published Safety Frameworks
On the governance side, the Report documents that 12 AI companies published or updated what are called Frontier AI Safety Frameworks in 2025. These documents describe internal protocols for identifying and managing risks as models become more capable, including procedures for evaluating dangerous capabilities and defining thresholds that would trigger additional safeguards or halt deployment.
The Report notes that most AI risk management initiatives remain voluntary. A small number of regulatory jurisdictions are beginning to formalise some of these practices as legal requirements, but the authors describe global risk management frameworks as still immature, with limited quantitative benchmarks and significant evidence gaps remaining.
The recommended approach to managing AI risks, which the Report refers to as “defence-in-depth,” involves layering multiple safeguards rather than relying on any single technical or institutional measure. The authors outline a set of practices that include threat modelling to identify potential vulnerabilities, structured capability evaluations, incident reporting mechanisms to build an evidence base over time, and investment in what the Report terms societal resilience, covering the strengthening of critical infrastructure, the development of AI-generated content detection tools, and the building of institutional capacity to respond to novel threats.
International Cooperation Context
The 2026 Report is the second in a series initiated following the AI Safety Summit at Bletchley Park in November 2023. Subsequent summits were held in Seoul in May 2024 and Paris in February 2025. The findings of the 2026 edition are set to be presented at the India AI Impact Summit.
The Expert Advisory Panel that guided the Report’s development included nominees from Australia, Brazil, Canada, Chile, China, France, Germany, India, Indonesia, Japan, Kenya, Nigeria, Rwanda, Saudi Arabia, Singapore, South Korea, Turkey, Ukraine, the United Arab Emirates, the United Kingdom and the United States, among others, as well as representatives from the EU, OECD and UN.
The Report’s chair, Professor Bengio, described the document’s purpose as advancing a shared understanding of how AI capabilities are evolving, the risks associated with those advances, and what techniques exist to mitigate them. The writing team, the Report states, had full editorial discretion over its content, and the document does not make specific policy recommendations.
The Report covers research published before December 2025. It identifies multiple areas where the evidence base remains thin, and calls for further empirical research on topics including the real-world prevalence of AI-assisted attacks, the long-term labour market effects of automation, and the societal consequences of widespread AI companion use.
Other People are Reading
Continue Reading
