Contribution
My Lords, I too rise to support Amendment 41 in the name of the noble Baroness, Lady Harding, and to support Amendments 65, 66 and 67, to which I have also put my name. As the noble Baroness, Lady Harding, said, to have a Bill with “cyber security” in the title and not mention AI seems to have been a huge error. The Government seem to have recognised the great lacuna in some of their amendments in this group, which is most welcome. Also, we must welcome the Chancellor of the Exchequer’s low-key announcement over the weekend that the Government would invest £120 million in the Artificial Intelligence Security Institute. Surely this Bill should be an opportunity to ensure that AISI investment is used as effectively as possible and that, just as this country recognises the importance of AI in keeping our technical leading role, this new, incredibly powerful technology has new frontier models that need to be investigated prior to release.
Amendment 41 is an important step forward in ensuring that AI developers notify the relevant authorities when things go wrong, so if the noble Baroness decides to take this amendment to the vote, I will support her. I am surprised that not all parties think that the measures in this amendment need to be supported. AI frontier models have proved to be a threat, and we need to face up to it. The AI providers’ own safety report tells us of the threats posed by new models. Google wrote that a person outside the company had attempted to use its AI tool, Gemini, to obtain
“a complete, step-by-step technical guide for synthesizing weaponized biological agents”.
An Anthropic report found that there were five case studies of actors using its models in a way that could support biological weapons development. There were also reports of the terrorist group, Boko Haram, using frontier AI, including ChatGPT, to plan attacks and design explosive devices. Surely these concerns should encourage the Government at least to look at my noble friend Lord Tarassenko’s Amendment 66, which would have safety against terrorists and biological attacks, among other things, built into new AI frontier models and put AISI in a stronger position to be a third-party evaluator of frontier models.
I had such hopes when the 2023 Bletchley Park summit announced that AISI and the model providers would work together and voluntarily submit their new models to the directorate, but this is decreasingly happening. Anthropic’s Claude Mythos 5.1 was withdrawn from AISI last month. There are other cases such as Claude Opus 5.5, which was reportedly withdrawn from AISI’s pre-release evaluation in September, as was the OpenAI and GPT-6 Luna model.
I hope that your Lordships will listen to the chair of the Business, Innovation, Science and Trade Committee in the other place, who this month during an evidence-gathering session with AI leaders announced that AISI lacked compulsory access to models and a power to stop new models coming to the market. I call on the Minister to address these concerns.
This brings me to Amendment 67, in the name of my noble friend Lady Kidron. As many other noble Lords have said during the passage of this Bill through the House, this is a defensive Bill. Surely the best way to stop new frontier models from going rogue is to be proactive in putting guardrails around their design. Otherwise, we are just going to play a game of whack-a-mole. The principles in the Online Safety Act were to try to instil safety by design into the algorithms of online content providers. Surely the same needs to happen to AI frontier models. Amendment 67 sets out to provide guidelines for how the design of AI models might be kept in check.
I ask His Majesty’s Government to ensure that safety is built into new AI model designs and that the highly respected AISI is given sufficient power to ensure that they are safe. The Minister must know that the world would be grateful if new models were safe by design and properly evaluated before release.