International AI Safety Report 2026

Summary

The International AI Safety Report 2026 is the most authoritative scientific assessment of AI safety to date. Commissioned after the 2023 Bletchley Park AI Safety Summit, the report assembles over 100 researchers and domain experts from more than 30 countries under the leadership of Professor Yoshua Bengio. Its mandate is straightforward but consequential: give governments an independent, evidence-based evaluation of where frontier AI capabilities stand, what risks they pose, and whether current safeguards are adequate.

What the Report Assesses

The report takes a structured approach to evaluating AI safety across three connected dimensions. First, it examines the current state of frontier AI capabilities, documenting how large-scale models have progressed in reasoning, code generation, scientific analysis, and autonomous action. Rather than speculating about distant futures, the assessment grounds itself in observed capabilities and near-term trajectories based on published benchmarks and disclosed system performance.

Second, the report catalogs the risks these capabilities introduce. It moves beyond abstract concerns to document specific failure modes, deployment incidents, and emerging threat patterns observed or credibly demonstrated in controlled settings.

Third, the report evaluates the effectiveness of existing mitigation strategies, from technical safety measures implemented by developers to governance frameworks adopted by governments. This gap analysis between current risks and current safeguards forms the core of the report's policy relevance.

The Three Risk Layers

The report organizes AI risks into three distinct but overlapping layers, each requiring different governance responses.

Malicious Use

The first layer addresses deliberate misuse of AI systems by bad actors. The report documents how frontier models can lower the barrier to entry for cyberattacks, disinformation campaigns, and the development of dangerous materials. It examines case studies where AI tools have generated convincing phishing content at scale, automated vulnerability discovery in software systems, and produced synthetic media designed to manipulate public discourse. The panel notes that while many of these risks existed before large-scale AI, the automation and accessibility that current models provide represent a qualitative shift in threat severity.

Malfunctions and Unintended Behavior

The second layer covers risks arising not from deliberate misuse but from systems behaving in ways their developers did not anticipate. This includes hallucination in high-stakes decision contexts, unexpected emergent behaviors in multi-agent deployments, and failures in alignment between stated objectives and actual system behavior. The report pays particular attention to the growing deployment of AI systems in critical infrastructure, healthcare, and financial services, where malfunctions can cause direct harm. Current evaluation methods remain insufficient for predicting failure modes in novel deployment contexts.

Systemic Risks

The third layer examines broader societal and economic effects that emerge from widespread AI adoption rather than from any single system or incident. These include labor market disruption at a pace that outstrips retraining capacity, concentration of economic power among a small number of AI developers, erosion of epistemic trust through synthetic content saturation, and the potential for competitive dynamics between nations to undermine safety standards. The report argues that systemic risks are the least well-governed category, partly because they cross jurisdictional boundaries and partly because they unfold gradually rather than through dramatic incidents.

Key Findings on Frontier AI

The report reaches several findings with direct implications for governance. The pace of capability improvement in frontier models continues to exceed the pace of safety research, creating a widening gap that current voluntary commitments have not closed. The report documents significant variation in safety practices across developers, with some organizations investing heavily in red-teaming, interpretability research, and deployment safeguards while others treat safety as a secondary concern.

Existing evaluation frameworks for AI safety are not keeping pace with the systems they are meant to assess. Benchmarks that were informative two years ago may no longer capture the relevant capability thresholds, and there is no internationally agreed-upon methodology for conducting safety evaluations of frontier models.

On the positive side, the report acknowledges meaningful progress in several areas: the establishment of AI safety institutes in multiple countries, more sophisticated red-teaming methodologies, and growing industry adoption of structured safety cases before deployment. However, these advances remain voluntary and unevenly distributed.

Why Governance Teams Should Read This

For organizations building or deploying AI systems, this report serves as a reference for several reasons. It provides the most current picture of what frontier AI can actually do, cutting through both hype and dismissiveness with empirical grounding. The three-layer risk framework offers a practical structure for organizing internal risk assessments, ensuring that governance programs address malicious use, malfunctions, and systemic effects rather than focusing narrowly on one category.

The report's gap analysis between risks and mitigations highlights where industry practices leave gaps of what the scientific community considers adequate. Governance teams can use these findings to benchmark their own safety programs and identify areas requiring additional investment. The international scope also makes it valuable for organizations operating across multiple jurisdictions, as it maps the emerging global consensus on which risks are most pressing and which governance responses are gaining traction.

This report also signals the direction of future regulation. Governments that commissioned the assessment are using its findings to inform policy decisions. Organizations that align their governance practices with the report's recommendations position themselves ahead of regulatory requirements rather than scrambling to catch up after rules are finalized.