OpenAI has slowed its model development process as cybersecurity risks linked to AI continue to rise. The company cited the potential for its upcoming Astra model to gain critical cyberattack capabilities as a key factor in the decision. OpenAI paused reinforcement learning for two weeks, with its 'largest planned frontier RL run' remaining on hold. Workloads that have not met new security requirements are also suspended. The Hugging Face security incident and 'rapid progress in our internal research' prompted the slowdown, according to the company. Since then, research environments have been hardened with better network isolation and stricter sandboxes. A new monitoring system alerts within 30 minutes of detecting suspicious behavior, using roughly 20 percent of supervised inference compute depending on workload. The company plans to expand its Preparedness Framework and invest more in alignment research. It has, however, disbanded the team behind that framework, shifting responsibilities to other teams.

Critics are likely to accuse OpenAI of fear-mongering to buy time and attention, but the independent government agency AISI has documented similar harmful model behavior, lending some weight to OpenAI's claims. The company emphasized that its actions are driven by the need to address real security concerns rather than public relations strategies. OpenAI has also highlighted the importance of maintaining strict security protocols to prevent potential misuse of AI technologies. The company said, 'We are prioritizing safety over speed to ensure our models do not pose a threat to cybersecurity.'

OpenAI has taken a proactive approach to addressing these risks, implementing new monitoring systems and tightening security measures across its research environments. The company's focus on safety has led to a reorganization of its teams, with responsibilities for the Preparedness Framework now handled by other departments. OpenAI said, 'Our goal is to ensure that our models are as secure as possible before they are released to the public.'

Source: thedecoder