Musk, Altman back proposal to slow frontier AI development

Musk endorsed the proposal in a post on X, saying Amodei was right.

Altman also expressed support on the platform, saying the pace of frontier AI development had become a major subject of discussion at OpenAI in recent weeks. He added that the company would follow Anthropic in granting independent evaluators ongoing access comparable to that of employees.

Amodei outlined his proposal in an essay published in September. He argued that the capabilities of advanced AI models were progressing faster than the mechanisms designed to monitor and control them.

The Anthropic chief stressed that his proposal did not call for an end to model training or technical progress. Instead, developers would take additional time to test safeguards and ensure that safety research keeps pace with expanding model capabilities.

His three-part plan calls for permanent access for independent evaluators at leading AI companies, common safety standards across the industry and international coordination on the development and use of powerful AI systems.

Anthropic has committed to implementing the first measure. External evaluators would examine its safety practices and training processes, report incidents and publish key findings without editorial control from the company, subject to limited legal, security and confidentiality restrictions.

Amodei said an additional year or two could help researchers improve model alignment, testing, interpretability and operational security.

He also warned that groups of increasingly capable autonomous agents could create serious cybersecurity risks. In his assessment, systems developed within the next six to twelve months could potentially conduct attacks through persistent botnets on a much larger scale if safeguards fail to keep pace.

The proposal follows a security incident involving AI agents developed by OpenAI. During internal testing, the agents bypassed parts of an isolated environment, accessed the internet and reached systems operated by OpenAI and Hugging Face.

OpenAI subsequently introduced tighter isolation for research workloads, additional network controls and expanded monitoring for unauthorized activity. The company also temporarily slowed parts of its model development program, including a two-week suspension of reinforcement learning training for models intended for release.

Earlier, Qazinform reported that an AI researcher who had spent the previous three years developing frontier models at OpenAI and Anthropic resigned, warning that leading AI companies were moving too quickly toward potentially uncontrollable superintelligent systems.