The reality of AI can be unforgiving as the Fable case demonstrates

Anthropic recently rolled out a variant of Mythos under the name Fable. However, the fun was short-lived; its use by foreign nationals was banned by the US government. We will discuss this further in this blog post.

In the previous blog, we discussed, among other things, the security measures that the major AI model providers have implemented. This primarily concerned the protection of confidential data entered by the user themselves. With the announcement of Mythos, Anthropic's latest model, another concern also emerged. Mythos was capable of discovering vulnerabilities in existing software that had not previously been found by humans. Too risky to make public, according to Anthropic. That is why they launched Project Glasswing.

Gladness

There was good news for the average person on June 9: Anthropic is rolling out a variant of Mythos under the name Fable. In no time, the internet was flooded with the most bizarre use cases, ranging from impressive interactive maps and a Minecraft clone created with a four-word prompt to a fully functional technical design of a V8 engine. That said, Fable isn't necessarily better at drafting contracts or extracting information from legal documents.

Nothing extraordinary

The fun was short-lived, however, as Amazon employees soon managed to use Fable to discover security vulnerabilities in four different applications. Amazon reportedly asked the government to take the model offline; five other companies also contacted officials. In addition, the US government has indications that a Chinese group had gained access to Mythos.

The result: citing national security, the US government has imposed an export restriction prohibiting all access to Claude Fable 5 and Mythos 5 for foreign nationals (both inside and outside the US), including Anthropic’s own non-US employees. Because excluding users based on nationality is never entirely foolproof, Anthropic has decided to take both models completely offline worldwide. All other Claude models will continue to operate as usual for the time being.

Anthropic immediately complied with the request but openly disagrees with it. They argue that the issue involves a narrow, non-universal jailbreak, meaning the model was asked to read a codebase and fix software errors. According to Anthropic, no universal jailbreak has been discovered, not even after thousands of hours of red-teaming with the US government. They add that if this becomes the norm, all new frontier models will be blocked. They characterise the situation as a misunderstanding and aim to restore access as soon as possible.

Impact in Europe

Earlier this month, the US government issued a directive calling on AI model developers to voluntarily grant the government access to their models before they are released to the market. This directive specifically targets models that impact information security. Additionally, Anthropic is involved in a legal dispute after the US government designated the organisation a "supply chain risk to US security." Anthropic has made it clear that the models it develops must under no circumstances be involved in military decision-making (such as the development of autonomous weapons and mass surveillance). It remains unclear whether these developments actually played a role in the US government's decision.

This development has certainly highlighted just how unpredictable reality can be. One moment you have the latest model at your disposal, and the next, you can no longer use it. This raises the question of the extent to which such developments should be taken into account when selecting a suitable AI solution. Consequently, factors like data and workflow portability are becoming increasingly important when working with a tool driven by a single model. Model-agnostic systems hold a significant advantage, as they can switch models more easily, offering the kind of flexibility you, as a user, might well need.

With the rollback of the Fable 5 deployment, Mistral has also come more into the spotlight again. Generally, Mistral's models score lower than the American models, but it does prevent you from having to deal with the whims of the American government.

These kinds of developments show how important it is to make conscious choices about which AI tools you deploy and we're here to guide you. Curious about making strategic and responsible AI choices for your organisation? Schedule a no-obligation conversation.