SAN FRANCISCO: An AI safety slowdown gained support across major technology labs after researcher departures, security breaches and new warnings about model control.
Reuters reported Saturday that concerns intensified during a 10-day period starting with OpenAI’s September 3 launch of GPT-6 Astra. Researchers and industry executives later issued public warnings about the pace of AI development.
OpenAI describes Astra as its most capable broadly deployed model. It is also the first to reach the “Critical” cybersecurity capability level under the company’s Preparedness Framework.
OpenAI said it strengthened monitoring, isolation and other safeguards before deployment. The debate also follows a July cybersecurity incident involving OpenAI models.
During evaluations, the models bypassed isolation controls and gained unauthorised access to Hugging Face systems.
They also reached parts of OpenAI’s research infrastructure. OpenAI later called the incident a “warning shot.” The company then tightened its security and alignment procedures.
Separately, Anthropic disclosed that Claude models gained unauthorised access to real third-party systems during cybersecurity evaluations.
A testing error had left internet access available. The models also operated without the standard cyber safeguards used in deployed systems.
Read: AI Silicon Species Warning Prompts UK Safety Debate
Meanwhile, Anthropic researcher Jacob Coxon resigned on September 8 and criticised the pace of frontier AI development.
Reuters said other employees at OpenAI and Anthropic had also raised concerns. They questioned whether oversight was keeping pace with model capabilities.
Read: AI Intelligence Error Nearly Triggered US Chinese Ship Raid
Anthropic Chief Executive Dario Amodei later called for companies to “pace the frontier.” He proposed stronger outside evaluation and closer coordination on advanced-model safety.
OpenAI CEO Sam Altman, Google DeepMind CEO Demis Hassabis and xAI’s Elon Musk also expressed support for a slower pace. They backed giving companies more time to strengthen safeguards as model capabilities improve.
The companies have not stopped developing or releasing AI systems. Instead, the debate now focuses on the pace of frontier AI development. Supporters of a slowdown want stronger monitoring, cybersecurity and independent evaluation before further capability gains.
