Major Safety Assessment Published

The International AI Safety Report 2026 was published on 3 February 2026, providing a comprehensive, science-based assessment of the capabilities and risks of general-purpose AI. The report is led by Turing Award winner Yoshua Bengio, backed by an Expert Advisory Panel with nominees from more than 30 countries and international organizations, and authored by over 100 AI experts.

Exponential Computing Growth

Computing power has grown exponentially since 2012, with the largest training runs likely surpassing 10^26 FLOP in 2025. This acceleration has enabled new model architectures and capabilities that present both opportunities and risks.

Reasoning Models Drive Performance Gains

Reasoning models—which work through problems by generating and comparing multiple outcomes before selecting a final answer—have become more common, leading to improved performance in mathematics, coding, and scientific tasks.

AI Agent Capabilities Remain Limited

AI agents have become a major focus of development. Although agents have demonstrated the ability to complete a variety of software engineering tasks with limited human oversight, they cannot yet complete the range of complex tasks and long-term planning required to fully automate many jobs.

Criminal Misuse and Deepfake Concerns

The report documents growing misuse of general-purpose AI’s ability to generate high-quality text, audio, images, and video for criminal purposes, including scams, fraud, blackmail, extortion, defamation, non-consensual intimate imagery, and child sexual abuse material.

AI-generated deepfakes are becoming more realistic and harder to identify. Personalised deepfake pornography disproportionately targets women and girls, representing a significant and emerging harm.

Influence Operations and Belief Manipulation

In experimental settings, AI-generated content can produce measurable changes in people’s beliefs. People interacting with content produced by models with more computing power are more likely to change their views.

State and Criminal Actor Involvement

Criminal groups and state-associated attackers are actively using general-purpose AI in their operations. AI currently plays its largest role in scaling the preparatory stages of cyberattacks.

Biological and Chemical Threats

General-purpose AI systems can help enable the creation of biological and chemical threats by producing laboratory instructions, helping troubleshoot experimental procedures, and answering technical questions.

Healthcare and AI Assistance Risks

A study found that clinicians’ rate of detecting tumors during colonoscopy was 6% lower after several months of performing the procedure with AI assistance, raising concerns about automation-induced complacency.

AI Companion Products and Psychological Risks

AI companion apps have grown to tens of millions of users. Studies have found some products may foster psychological dependence, reinforce harmful beliefs, or encourage dangerous actions.

Safeguard Resilience Challenges

Although developers have made it more difficult to bypass model safeguards, new attack techniques are constantly being developed, and attackers still succeed at a moderately high rate.

Building safer models is inherently difficult because there is no universal consensus on what constitutes desirable AI behaviour.

Open-Weight Models and Monitoring Gaps

Open-weight models facilitate research and innovation, but their safeguards can be more easily removed and monitoring use is more challenging because anyone can run them outside of controlled environments.

Industry Safety Framework Progress

12 companies published or updated Frontier AI Safety Frameworks in 2025, which describe how companies plan to manage risks as they build more capable models.

OpenAI and Anthropic Safety Evaluation

OpenAI and Anthropic completed a first-of-its-kind joint safety evaluation on 27 August 2025, testing each other’s publicly released models for misalignment, instruction following, hallucinations, jailbreaking, and other safety concerns.

On jailbreaking evaluations, Claude models performed less well compared to OpenAI o3 and o4-mini, with Claude models with reasoning disabled outperforming Claude with reasoning enabled in one jailbreak scenario.

On hallucination evaluations, Claude models had an extremely high rate of refusals—as much as 70%—showing these models are aware of their uncertainty, while OpenAI o3 and o4-mini show lower refusal rates with higher hallucination rates.

OpenAI o3 showed robustness across a range of challenging misalignment and safety evaluation scenarios in Anthropic’s tests, with reasoning models tending to give the strongest performance across multiple evaluation categories.

GPT-5 Launch

OpenAI launched GPT-5 in early August 2025, a unified system of models that brings the benefits of reasoning models to all users, showing substantial improvements in areas like sycophancy, hallucination, and misuse resistance.

Source: Inside Privacy