We deployed open-source AI agent OpenClaw inside a sandbox on our research platform. Despite our initial countermeasures, it successfully identified our organisation by name, inferred the identity of a human operator and reconstructed a timeline of some of our research activities.

We conducted cyber evaluations of Anthropic’s Claude Mythos Preview and found continued improvement in capture-the-flag (CTF) challenges and significant improvement on multi-step cyber-attack simulations.

Sharing work with the National Cyber Security Centre (NCSC) on how cyber defenders can use advanced AI capabilities to stay ahead of attackers.

A monitoring method and large‑scale analysis to understand the tasks AI agents are performing today.

We introduce SandboxEscapeBench, the first benchmark to systematically evaluate whether AI agents can break out of their sandboxes, and share some early results.

We tested seven large language models (LLMs) on two custom-built cyber ranges, measuring their ability to execute extended attack sequences in complex environments.

Alongside Irregular, we found evidence demonstrating that evaluators need to use large token budgets to understand the cyber capabilities of recent Large Language Models (LLMs).

We developed a scalable approach to measuring how text-based AI models can assist in three complex fraud and cybercrime scenarios.

We outline a step-by-step pipeline for using our open-source transcript analysis tool, Inspect Scout.

The Alignment Project welcomes its first cohort of grantees, and new partners join the coalition, bringing total funding to £27m.

New partnership exploring the security and societal implications of voice AI systems

Introducing an automated attack technique that generates universal jailbreaks against the best defended systems

The International Network for Advanced AI Measurement, Evaluation and Science reflects on recent meeting and looks ahead to the India AI Impact Summit

Alongside the government’s new Future of Work Unit, we conducted a pilot study to explore how much AI models increase worker productivity for common tasks.

Adam Beaumont, Director of the UK AI Security Institute, reflects on the year's biggest achievements.

Our inaugural Frontier AI Trends Report draws on 2 years' worth of evaluations to provide accessible insights into the trajectory of AI development.

How we’re partnering with government and experts to prevent the creation and spread of AI‑generated CSAM

Our new paper shares findings from an adversarial evaluation of monitoring systems for detecting sabotage by AI coding agents.

Expanding our collaboration with a new research MOU

Our new paper shares the results of an auditing game to evaluate ten methods for sandbagging detection in AI models.

A deep dive into AISI’s study of the persuasive capabilities of conversational AI, published today in Science.

Insights from our alignment evaluations of Claude Opus 4.1, Sonnet 4.5, and a pre‑release snapshot of Opus 4.5.

An overview of the research we’ll be presenting at this year’s NeurIPS conference.

Takeaways from expert interviews on barriers to AI capable of automating most cognitive labour.

Our dedicated library to make AI control experiments easy, consistent, and repeatable.

Why we use transcript analysis for our agent evaluations, and results from an early case study.

Our work with Anthropic and the Alan Turing Institute suggests that data poisoning attacks may be easier than previously believed.

What we learned from a large-scale empirical study of AI use for political information-seeking.

Insights into our ongoing voluntary collaborations with Anthropic and OpenAI.

Exploring how far cyber security approaches can help mitigate risks in generative AI systems, in collaboration with the National Cyber Security Centre (NCSC).

Current methods and open problems in open-weight model risk management.

A comprehensive toolkit for safely evaluating AI agents.

We outline our approach to study and address AI risks in real-world applications

Advancing methodologies for agentic evaluations across domains, including leakage of sensitive Information, fraud and cybersecurity threats.

Calibrating AI risk assessment through rigorous elicitation practices.

An introduction to white box control, and an update on our research so far.

Our new framework can assess the reliability of LLM evaluators, while simultaneously answering a primary research question.

How we're working to track and mitigate against criminal misuse of AI.

An Example Safety Case for Safeguards Against Misuse

HiBayES: a flexible, robust statistical modelling framework that accounts for the nuances and hierarchical structure of advanced evaluations.

We outline our research priorities, our approach to developing technical solutions to the most pressing AI concerns, and the key risks that must be addressed as AI capabilities advance.

A comprehensive benchmark to detect emerging replication abilities in AI systems and provide a quantifiable understanding of potential risks

Our new paper outlines how AI control methods can mitigate misalignment risks as capabilities of AI systems increase

20 Systemic Safety Grant Awardees Announced

We outline our approach to technical solutions for misuse and loss of control.

Our new papers show how safety cases can help AI developers turn plans in their safety frameworks into action

Our new paper proposes core principles for evaluating misuse safeguards

The UK Artificial Intelligence Safety Institute and the U.S. Artificial Intelligence Safety Institute conducted a joint pre-deployment evaluation of OpenAI's o1 model

A Methodology for Evaluating Scientific Assistants

The UK Artificial Intelligence Safety Institute and U.S. Artificial Intelligence Safety Institute conducted a joint pre-deployment evaluation of Anthropic’s latest model

How to write part of a safety case showing a system does not have offensive cyber capabilities

The AI Safety Institute reflects on its first year

We’re open-sourcing dozens of LLM evaluations to advance safety research in the field

We are launching a bounty for novel evaluations and agent scaffolds to help assess dangerous capabilities in frontier AI systems.

We look into the evolving role of third-party evaluators in assessing AI safety, and explore how to design robust, impactful testing frameworks.

Calling researchers from academia, industry, and civil society to apply for up to £200,000 of funding.

Our Chief Scientist, Geoffrey Irving, on why he joined the UK AI Safety Institute and why he thinks other technical folk should too

We studied whether people want AI to be more human-like.

A common technique for quickly assessing AI capabilities is prompting models to answer hundreds of questions, then automatically scoring the answers. We share insights from months of using this method.

AISI is bringing together AI companies and researchers for an invite-only conference to accelerate the design and implementation of frontier AI safety frameworks. This post shares the call for submissions that we sent to conference attendees.

AISI funded Epoch AI to explore AI researchers’ differing predictions on the automation of AI research and development and their suggestions for how to evaluate relevant capabilities.

As a complement to our empirical evaluations of frontier AI models, AISI is planning a series of collaborations and research projects sketching safety cases for more advanced models than exist today, focusing on risks from loss of control and autonomy. By a safety case, we mean a structured argument that an AI system is safe within a particular training or deployment context.

We are opening an office in San Francisco! This will enable us to hire more top talent, collaborate closely with the US AI Safety Institute and engage even more with the wider AI research community.

Since February, we released our first technical blog post, published the International Scientific Report on the Safety of Advanced AI, open-sourced our testing platform Inspect, announced our San Francisco office, announced a partnership with the Canadian AI Safety Institute, grew our technical team to >30 researchers and appointed Jade Leung as our Chief Technology Officer.

We tested leading AI models for cyber, chemical, biological, and agent capabilities and safeguards effectiveness. Our first technical blog post shares a snapshot of our methods and results.

This is an up-to-date, evidence-based report on the science of advanced AI safety. It highlights findings about AI progress, risks, and areas of disagreement in the field. The report is chaired by Yoshua Bengio and coordinated by AISI.

We open-sourced our framework for large language model evaluation, which provides facilities for prompt engineering, tool usage, multi-turn dialogue, and model-graded evaluations.

The UK and US AI Safety Institutes signed a landmark agreement to jointly test advanced AI models, share research insights, share model access and enable expert talent transfers.

The UK AI Safety Institute and France’s Inria (The National Institute for Research in Digital Science and Technology) are partnering to advance AI safety research.

This post offers an overview of why we are doing this work, what we are testing for, how we select models, our recent demonstrations and some plans for our future work.

Since October, we have recruited leaders from DeepMind and Oxford, onboarded 23 new researchers, published the principles behind the International Scientific Report on Advanced AI Safety, and began pre-deployment testing of advanced AI systems.

At the first AI Safety Summit at Bletchley Park, world leaders and top companies agreed on the significance of advanced AI risks and the importance of testing.

Since September, we have recruited leaders from OpenAI and Humane Intelligence, tripled the capacity of our research team, announced 6 new research partnerships, and helped establish the UK’s fastest supercomputer.

In our first 11 weeks, we have recruited an advisory board of national security and ML leaders, including Yoshua Bengio, recruited top professors from Cambridge and Oxford and announced 4 research partnerships.