Donny Greenberg, an expert in training artificial intelligence models at Anthropic, welcomed the “constructive” dialogue initiated with the European Commission and the European Union Agency for Cybersecurity (ENISA) regarding access to its frontier AI model, Mythos, while calling for global cooperation to address the cybersecurity risks linked to these models, taking the view that no organisation and no country can tackle them alone.
“Our goal is to enable these models to eliminate the risks of cyberattacks targeting all critical infrastructure: financial, physical, civic and digital infrastructure”, he said on Tuesday 14 July, during an exchange within the European Parliament’s Committee on the Internal Market and Consumer Protection (IMCO), while meanwhile leaving the majority of MEPs’ questions unanswered.
MEPs asked him whether the company believed that the European Union should develop its own alternatives and whether Anthropic could help Europe design its own models. They also asked who ultimately controlled the ‘kill switch’ enabling the models to be deactivated – the US government or the company? – and whether Anthropic could confirm who held European citizens’ data and whether that data would be accessible to the United States government.
Mr Greenberg stressed that it was important not to look at “individual models in isolation”, but to make comparisons “between different models” in order to identify “the weakest link” in terms of vulnerability across different sectors and software, which requires “close cooperation and coordination (…). This is a global problem”. Anthropic expects that “many other AI companies will have models in this category, and that some will place them on the market without security safeguards” over the next three to six months.
“We are working tirelessly to avoid this kind of complex gradual deployment in future”, he assured, in response to questions on the export restrictions imposed by the US government on Fable and Mythos models by Anthropic for non-US citizens, restrictions that were lifted in July. He acknowledged that, in the specific case of Mythos, “no particular modification” had been made to the model in order to lift the restrictions. (Original version in French by Ana Pisonero Hernández)