Pioneering Government Institute for AI Security to Open in San Francisco

Pioneering British governments AI The Safety Institute is set to expand its international horizons by opening its first overseas office in San Francisco this summer, Technology Secretary Michelle Donelan announced today (Monday, May 20).

This expansion marks a crucial step that will allow the UK to harness the wealth of technology talent available in the Bay Area, to engage with the world's largest players. AI laboratories based in London and San Francisco, and cement relationships with the United States to advance AI security for the public interest.

The office should open its doors this summer and recruit a first technical team led by a research director.

This will be a complementary branch to the Institute's headquarters in London, which continues to grow and already has a team of over 30 people. The London office will continue to grow and gain the expertise needed to assess border risks. AI systems.

By expanding its presence in the United States, the Institute will establish close collaboration with the United States, thereby strengthening the country's strategic partnership and approach to AI security, while sharing research and conducting joint assessments of AI models that can shed light AI security policy around the world.

Secretary of State for Science and Technology Michelle Donelan said:

This expansion represents British leadership in AI in action. This is a pivotal moment in the UK's ability to explore both the risks and potential of AI from a global perspective, strengthening our partnership with the United States and paving the way for other countries to leverage our expertise as we continue to lead the world in this area. AI security.

Since the Prime Minister and I founded the AI Safety Institute, it has gone from strength to strength and in just over a year, here in London, we have built the first world government AI research team, attracting top talent from across the UK and beyond.

Opening our doors abroad and building on our alliance with the United States is at the heart of my project to set new international standards in AI security which we will discuss at the Seoul summit this week.

The expansion comes as the U.K. AI The Safety Institute publishes a selection of recent safety testing results from five publicly available advanced tests. AI models: the first government-backed organization in the world to reveal the results of its assessments.

Although only a small part of the Institute's broader approach, the results show the significant progress the Institute has made since November. AI Safety Summit as it strengthens its cutting-edge safety testing capabilities.

The Institute assessed AI models against four key risk areas, including the actual effectiveness of protections installed by developers in practice. As part of the results, the Institute's tests revealed that:

• Several models have addressed cybersecurity challenges, while struggling to meet more advanced challenges.
• Several models demonstrate similarities with PhD-level knowledge in chemistry and biology.
• All models tested remain highly vulnerable to basic jailbreaks, and some will produce harmful results even without dedicated attempts to bypass protections.
• The models tested were unable to complete more complex and time-consuming tasks without human supervision.

AI Safety Institute President Ian Hogarth said:

The results of these tests mark the first time we have been able to share some details of our model evaluation work with the public. Our evaluations will contribute to an empirical assessment of the model's capabilities and the lack of robustness of existing safeguards.

AI security is still a very young and emerging field. These results represent only a small part of the evaluation process AISI expands. Our ambition is to continue to push the boundaries of this field by developing cutting-edge assessments, with a focus on national security risks.

AI Safety remains a key priority for the UK as it continues to advance the global debate on the safe development of technology.

This effort was launched in November AI Security Summit at Bletchley Park, and momentum continues to build as the UK and Republic of Korea prepare to co-host the AI Seoul summit this week.

As the world prepares to gather in Seoul this week, the UK is committed to working with Canada, including through their respective relationships. AI Security Institutes, to advance their ambition to create a growing network of state-backed organizations focused on AI security and governance. Confirmed by UK Technology Minister Michelle Donelan and Canadian Science and Innovation Minister François-Philippe Champagne, the partnership will serve to deepen existing ties between the two countries and inspire collaborative work on security research systemic.

Under the agreement, countries will work to share expertise to strengthen existing testing and evaluation work. The partnership will also enable secondment routes between the two countries and jointly identify areas of research collaboration.

Notes to editors

The Institute's security tests were carried out this year on five major publicly available language models (LLM) which are trained on large amounts of data. The models tested were anonymized.

The results only provide an overview of the model's capabilities and do not designate systems as secure or not. The tests that have been carried out represent a small part of the evaluation techniques AISI is currently developing and using, as outlined in the Institute's approach to evaluation, published earlier this year.

Today's post can be viewed at AI Security Institute website.

Today also marks the latest progress update from Institute President Ian Hogarth, who can be found here on the AI Security Institute website.