Last month, we hosted the second dinner in our AI Enlightenment series. It was last summer when Darren first came to me with the idea of creating a campaign and a series of events around artificial intelligence, but grounded in the values of the Scottish Enlightenment. While it has been a few months in the making, the first events we have hosted have been rooted in that spirit of collaboration and support.
We started these events with a very clear intention: we wanted to bring people together from entirely different organisations and disciplines to discuss, challenge, and support one another. Navigating business innovation and the rapid pace of development in the world of technology can feel isolating, and we wanted to change that.
Our initial discussions for the evening focused on the security risks of AI. We had all read the headlines on Anthropic's pre-release and the warnings about their new model, Mythos, and how powerful and advanced it was, outperforming other models across knowledge, reasoning, visual capabilities, and many other areas.
The security concerns focused mainly on its enhanced, larger context windows, which resulted in its ability to chain together minor vulnerabilities in systems. It works just like a human hacker would, connecting individual flaws to find a way in. Anthropic pre-released the model under the code name Project Glasswing to allow for the patching and fixing of years-old vulnerabilities that had previously gone undiscovered.
More recently, with the release of Fable 5, Anthropic's safe version of Mythos for public use, there has been far more news. Firstly, we saw sudden shifts in terms of service and a distinct lack of transparency. The fact that the model now retains data for 30 days, which causes huge data compliance issues for businesses, combined with the fact that it can silently downgrade your model or ‘nerf’ your prompts behind the scenes to restrict responses, has caused widespread conversations about data privacy, censorship, and the ethics of what Anthropic is trying to do in the market.
On top of that, the US government ordered Anthropic to limit foreign access, which led them to disable these most advanced models. This has raised very real, tangible questions about sovereignty, control, power, and access, not just at a national level, but also for businesses making critical, everyday decisions on AI.
All of these developments have massive implications for businesses, particularly when it comes to data privacy, ownership, and control. It represents a significant risk if your core business systems and daily processes are built on tools where the provider can suddenly change the rules, restrict your access, or downgrade performance without warning. And that is before we even touch upon the broader questions of intellectual property ownership and what data these tech companies are using to train their models.
We have already seen real-world examples of AI companies launching services that directly compete with their own partners. Recently, Anthropic's Chief Product Officer, Mike Krieger, resigned from Figma's board of directors. Just days later, Anthropic launched Claude Design, a tool capable of generating prototypes and user interfaces directly from text prompts. This sudden transition from partner to direct competitor caused Figma's stock to drop.
This raises a vital question for any leader: if you share your proprietary data, your unique ways of working, and your competitive edge with these external tools, what is to stop that knowledge from leaking out or being used by future competitors?
During our dinner, this specific worry turned our conversation toward open-source models, building proprietary systems, ensuring true data ownership, and exploring decoupled architectures, often called headless AI.
The main takeaway from our evening was simple: because this space is evolving so quickly, absolutely nobody has it all figured out. There are major questions we all need to ask about how we build our technology, how we structure our systems, and which external tools we choose to rely on. If you use a software-as-a-service product with built-in AI features, but you have no control over which model it runs, how your data is handled, or how the outputs are generated, how secure can your core business really be?
While earlier in the year only a few voices were raising concerns about sovereignty and control, those questions have now become unavoidable.
The risks to security and compliance are very real, but, as in previous eras of rapid innovation, these challenges also bring immense opportunity. The key is to approach them with thoughtful, strategic design, ensuring that you maintain ownership of your systems, your data, and ultimately, your success.
We covered so much ground over the course of the evening, happily wandering down rabbit holes and going off on tangents. We are incredibly passionate about keeping these conversations going, supporting open dialogue, and sharing what actually works. Here in Scotland, we have a uniquely supportive network and an amazing history of leading the way in global innovation. We will continue hosting these evenings, build on them, and collaborate with businesses across the country.
If you would like to join our next discussion, co-host an evening, or contribute to the conversation, please get in touch with Georgina. She supports our wider team to make these dinners happen, working across our client base and our broader network to bring people together and spark genuine collaboration.