The European Parliament recently held its first dedicated hearing focused on risks associated with advanced artificial intelligence. The result? Deep disappointment among lawmakers. The problem was not just the technology itself, but also the way Anthropic approached the hearing. Instead of the expected Head of Public Policy, it sent a technical expert, raising questions about its genuine willingness to engage in dialogue with regulators.

Cyber duality: The Mythos and Fable models

At the center of attention are two specific models from Anthropic – Mythos and Fable. Unlike the commonly available versions of Claude that we use for writing emails or analyzing data, these models demonstrate the ability to identify and exploit software vulnerabilities. In expert circles, this is referred to as dual-use AI. This means that the same tool can be used to strengthen cyber defense, but also for automated and devastating attacks on critical infrastructure.

European institutions are trying to gain access to these models to understand their potential for misuse. The concern is clear: if these capabilities are not under strict oversight, they could threaten the stability of digital systems across the entire EU. This is a topic that directly affects the Czech environment as well, especially in the context of protecting our energy grids and state administration from cyber threats.

For comparison, while models like GPT-4o from OpenAI or Gemini from Google primarily focus on general intelligence and multimodal interaction, Anthropic, within its research directions (including Mythos), explicitly concentrates on deeper security aspects and model capability control. This makes them a unique but also controversial player on the market.

Governance as a safeguard: The role of Ben Bernanke

Anthropic is aware, however, that without the trust of regulators, growth will not be sustainable in the long term. The company is trying to build institutional credibility through its Long-Term Benefit Trust. A recent and very significant step was the appointment of former U.S. Federal Reserve Chairman Ben Bernanke to this oversight body. Bernanke's presence is meant to signal that AI safety decisions will not be subordinated solely to profit motives but will be guided by independent experts.

This governance model is particularly interesting compared to the competition. While OpenAI has faced internal conflicts in the past regarding the balance between its nonprofit mission and commercial growth, Anthropic is trying to structurally protect its public mission through this Trust, which even has the authority to appoint or remove the majority of the board of directors. For both investors and regulators, this is an experimental model, but in the context of the upcoming EU AI Act, it represents an attempt at self-regulation that could serve as a template.

What does this mean for Czech companies and users?

Even though the incident took place in Brussels, the impacts will be felt by all of us. For Czech companies implementing AI into their processes, this has three main dimensions:

• Regulation: The implementation of the EU AI Act will require companies to maintain clear records of which models they use and whether they are "high-risk."
• Security standards: If models like Mythos prove to be truly risky, even stricter controls on technology exports and API access will follow.
• Tool availability: Claude models (including their latest iterations) are available in the Czech Republic through both the web interface and API. Prices are typically around $20/month for the Pro version, which is a comparable benchmark with ChatGPT Plus. Czech language support in Claude models is at a very high level, making them a preferred choice for local administrative and analytical tasks.

In the Czech context, it is also important to monitor the activities of the ENISA agency (EU Agency for Cybersecurity), which is already discussing cyber threat resilience with Anthropic. For Czech cybersecurity experts, this means that AI will not just be a tool for automation, but also a key element in defending against new types of attacks generated precisely by these advanced models.