On June 12th, the theoretical risk of AI took a real world form. With no notice, the US government flipped the kill switch on Anthropic’s Mythos model. The decision effectively cut off all access to the AI platform, not just for non-US companies, but across the world. Within 2 hours of the government decision, the tool was yanked offline, leaving developers and cyber security experts to wonder how to move forward. What questions should you be asking?
What happened?
You have probably used Anthropic services already this week, even if you were not aware. Anthropic Claude is one of the main foundation AI models. There are only a handful of these Large Language Models in the world at the moment. All other tools are built off these models. Open AI may feature in the news more, but Anthropic Claude is the foundation model of choice for Salesforce, JP Morgan, Zoom and Instaclass, and many other tools familiar around the office. It is better at complex analysis, deep coding and some more specialised workflows.
Anthropic developed two new sub-models, Mythos 5 and Fable 5 to be exceptionally good at coding. They were so good that they were finding thousands and thousands of vulnerabilities, like software bugs, across vital platforms that had passed all the existing benchmarking tests and were considered safe. The results of the latest 5 series were a huge concern to anyone in the industry, not least because a tool that can find vulnerabilities to protect people can easily be weaponised by bad actors to exploit those vulnerabilities. Fearing that the technology could be used to bypass all existing safety guardrails, the US government ordered Anthropic to halt all ‘exports’ of the technology with immediate effect.
What does this mean for businesses?
National security matters, and in the cyber security sector, there was a mix of fear about what harm Mythos could do, and celebration that it was genuinely helping to tighten up security.
Business leaders should be equally conflicted. AI is delivering the promised efficiencies that we have been pursuing for decades. But business leaders also love predictability. This event has proven to all that an enterprise’s core AI infrastructure can vanish over a weekend due to a political decision. That is not an environment business leaders want to operate in.
CEOs are currently being told to embed AI deep into their workflows in customer services, software development, and business operations. What happens if the tool is switched off today? Could your daily business operations still run?
The Digital Iron Curtain
The decision raises many questions, not least because the language of the order was termed as a ban on ‘exports’. Historically, export controls were for physical things like tanks, stealth tech and advanced semiconductor chips in wartime. Nvidia’s chips have been the centerpiece of much debate on who should access what in recent years.
This event is historic. This is the first time we have seen wartime-style export controls on digital products and code. But it didn’t just impact the ‘Rest of the World’. US workers were shocked to discover they were affected by the decision, too.
‘Banning exports’ was incredibly difficult for Anthropic to implement. They had no means to instantly verify the nationality of every single user or employee globally. The order also banned foreign nationals working on the tools inside the USA. Tech companies are built on the world’s genius. Anthropic did not have the means to instigate a partial shutdown and comply with the law, so they had no choice but a total shutdown for everyone. The 20th century concept of a physical border just does not map onto the 21st century global cloud infrastructure.
Politics vs AI safety
Support for regulation of AI remains high across almost all sectors of society. At the Institute of Analytics, we advocate for kill switches as part of a comprehensive safety approach. It was unclear how this decision fits into a pro-active model of systemic AI safety. It has done nothing to stop adversaries from building their own open-source frameworks, or from using alternative models. They have been shown the potential of AI for harm and are already building attack systems. Interestingly, the government’s decision to discuss the safety of Mythos was triggered by a request from the CEO of Amazon, a company that is building their own rival model.
It has sent shockwaves globally. European leaders have already identified the need for digital sovereignty, but are currently very limited in LLM options. Other regions share their concerns. If they pull away from American models to mitigate their risks, they could look to Chinese alternatives, like the highly affordable Zhipu, which is gaining in popularity. That just brings a different set of geo-political choices, rather than a solution.
What you can do as a business leader.
The Anthropic incident is a wakeup call that AI policy is increasingly becoming hard to distinguish from geopolitics. It proves what we have long argued: AI is not a standard Software as a Service utility, like Slack or Salesforce. It is a volatile and politically sensitive asset. Relying on any single, proprietary model, particularly a foreign one, introduces a single point of failure, and that brings risk.
One option in theory is to consider multi-model strategies, but that is not practical. In reality, in many countries, there are no regional providers of LLM foundation models. Across the whole of the African continent, there is one in South Africa, Vila. In Europe, GPT-NL is a sovereign model funded by the Dutch government. Outside of the US and China, there is very little provision.
An alternative would be to seek out open-weight models. These are smaller models that can be fine-tuned and then hosted locally to ensure that control of the business stays under the democratic control of the host country. There are more options once you look for regulatory-compliant open weight models, and they cannot be remotely deactivated. In Europe, these include in France (Mistral), Germany Mistral, (Aleph Alpha) and the Nordic countries (Silo AI). The Middle East has been spending heavily to position itself as a neutral global AI hub, with Falcon and Inception options both based in the UAE. There are similar smaller models spread across the Asia Pacific region.
The UK approach
The UK has taken a different approach. They recognise that ‘data residency’ is an illusion. It might be your data, subject to your sovereign laws of privacy. If you host an American model in a UK data centre, though, you still can’t guarantee that you can independently audit it, change its code or stop someone from hitting the delete switch on it. The real solution is to relocate the hardware back home. The UK has invested heavily in funding innovation in domestic hardware, and is developing open-source, locally hostable models through the Science of Fundamental AI Research Lab (SOFAIR).
The SOFAIR project, like many UK AI initiatives, is led by Higher Education. The University College London, Oxford, Cambridge and Edinburgh are working with the specific goal of building a new generation of highly efficient open-source foundation models that can run on standard hardware, rather than rely on massive foreign cloud data centres. The ambition is supported by UK-based cutting edge chip design, like Fractile and Olix, which are both trying to break Nvidia’s dominance. Arm Holdings’ groundbreaking semiconductor technology licenses their energy -efficient CPU and GPU designs to companies like Apple and Qualcom. Their technologies power over 99% of the world’s smartphones and home and wearable devices. With those developments in hardware, the 2030 targets for greater independence look a little more achievable. When those technologies are applied to AI, it may make high-performance compute systems more attainable for other countries.
A few years ago, I found myself trying to explain to governments that AI is not easy to regulate. We have extremely complex supply chains that cross many international borders. Governments have got the message, but businesses also need to come to terms with the fragile nature of the systems that they are becoming dependent on. The unthinkable - ‘Nobody would switch off a foundation model for political reasons’ - now has a case study,and it is not a theoretical business risk.
For many smaller companies, this simply means ensuring that the humans can still take over if the AI system stops - business continuity is protected. Larger companies need to treat AI as an international expansion decision. They are used to carrying out a risk report when expanding physically - opening up factories or retail outlets in a new country. It would be wise to treat any AI investment in a similar light and ask questions about where the model is hosted and whether those countries offer the political stability and stable relations to allow you to expect continuity.
IoA Members have the training in data and AI to deliver professional results. If learning to develop with Anthropic Claude is on your bucket list, or if you just want to learn more about how to work with AI, we have training ans support to get you there.
To learn more about how you can join the IoA, please visit our website on ioaglobal.org.uk. The Institute of Analytics are the leaders in professional ethical data standards and offer professional recognition, ongoing industry training, career development pathways, thought leadership to professionals in data roles and non-data roles..
