Anthropic's decision to skip UK pre-release tests for its Mythos 5.1 model has raised eyebrows, with the Financial Times reporting that this is the first time a major model has been kept from the UK's AI Security Institute (AISI) before launch. The move has sparked concerns about protectionism among American AI companies, with UK officials fearing a wider shift in this direction.

According to the FT, Anthropic did provide comparable US organizations with access to the model, fueling speculation about a potential protectionist stance. The UK government has not publicly criticized Anthropic, but a Cabinet Office spokesperson emphasized the importance of international cooperation in addressing AI-related risks. AISI has previously tested Anthropic models, including an evaluation of the Mythos preview's cyber capabilities in April.

The institute's testers found that the model was the first to complete a 32-step simulated network attack, doing so in three out of ten attempts. However, the decision to withhold the Mythos 5.1 model has been seen as a worrying precedent by some, including Miles Brundage, a former OpenAI policy researcher. Brundage argued that AISI has more overall capacity than its US counterpart, CAISI, and that the decision could undermine safety testing efforts.

Others, such as lawyer Preston Byrne, have suggested that the UK's approach to online speech regulation may be the root cause of the issue. The UK government's response has been measured, with a spokesperson highlighting the importance of continued cooperation between AISI and industry partners. The EU's cybersecurity agency, ENISA, has begun testing the Mythos 5 model, although it still does not have access to the updated Mythos 5.1 version.

The incident has sparked a debate about the role of independent safety testing in the development of AI models. Ed Newton-Rex, a campaigner on AI and creators' rights, has argued that the episode highlights the need for mandatory safety testing, rather than relying on labs to volunteer their models. As the AI landscape continues to evolve, the question of how to balance safety concerns with the need for innovation and cooperation will remain a pressing issue.

Este artículo fue escrito con la asistencia de IA.
News Factory APP - noticias agénticas para impulsar tu SEO y AEO.