Leading artificial intelligence companies are calling on the UK government to reveal more details about their AI security testing procedures. This highlights the potential limitations of relying solely on voluntary agreements to regulate this rapidly evolving technology.

This comes after OpenAI, Google DeepMind, Microsoft and Meta signed voluntary commitments in November 2023 to open their latest generative AI models for pre-deployment testing by the newly formed AI Safety Institute (AISI). While these companies agree to adjust their models based on identified deficiencies, they are now seeking clarity on the specific tests performed, their duration, and the feedback process should a risk be identified.

Although the government is not legally required to take action based on the test results, it emphasized its expectation that companies address identified risks. This highlights the ongoing debate about the effectiveness of voluntary agreements in ensuring the responsible development of AI. In response to these concerns, the government outlined plans for “future mandatory requirements” for leading AI developers to strengthen accountability measures.

Founded by Prime Minister Rishi Sunak, AISI plays a central role in the government's drive to position the UK as a leader in tackling the potential risks associated with the advancement of AI. These include the misuse of cyber attacks, the development of biological weapons and other harmful applications.

According to sources familiar with the process, AISI has already begun testing existing and unreleased models, including Google's Gemini Ultra. It focuses on potential risks of abuse, particularly cyber security, drawing on the expertise of GCHQ's National Cyber ​​Security Centre.

Government contracts reveal investments in capabilities to test for vulnerabilities such as jailbreaking (which causes AI chatbots to bypass security) and spear phishing (targeted email attacks). Efforts are also underway to develop “reverse engineering automation” tools for in-depth model analysis.

We will share findings with developers as appropriate. However, if risks are identified, we expect you to take all appropriate measures before implementing them.

The British government announced this. Financial times.

Google DeepMind acknowledged its collaboration with AISI and emphasized its commitment to working together to produce informed assessments and establish best practices for the evolving AI sector. OpenAI and Meta declined to comment.

In simple words:

Leading AI companies such as Google and Facebook are working with the UK government to safely test their new AI models before launching them. The government wants to ensure that these models are not used for hacking or producing dangerous weapons.

Companies will be happy to help, but would like more details about the tests, such as how long they last and what happens if problems are found. The government is trying to balance encouraging innovation with keeping people safe.