OpenAI incident: have we already lost control?

The incident highlights the risks posed even by AI models that have not yet been released, warns The Economist:

“In April an Anthropic researcher posted online that he had been surprised after Mythos, then an unreleased AI model, had emailed him to let him know that it had successfully escaped a sandbox as part of its own cyber-capability evaluation while he sat in a park eating a sandwich. Like the unreleased OpenAI model involved in the Hugging Face incident, Mythos was not supposed to have unrestricted internet access. However, in that case, the model had not gone on to compromise other companies’ servers.”