OpenAI is slowing some frontier model-development work as it reassesses cybersecurity risk, according to The Decoder. The report says the company describes the move as “pacing AI model development,” partly because an upcoming model called Astra may be approaching what OpenAI considers critical cyberattack capabilities. The operational changes are more concrete than a broad cautionary statement. The Decoder reports that OpenAI paused reinforcement learning for two weeks, that its “largest planned frontier RL run” remains on hold, and that workloads that have not met new security requirements are suspended. Those details suggest the company is applying the cyber-risk review directly to research operations rather than limiting it to policy language. The Decoder says OpenAI linked the slowdown to two factors: the Hugging Face security incident and rapid progress in its own internal research. The provided material does not specify the details of the Hugging Face incident, nor does it give technical benchmarks for Astra’s suspected cyber capabilities, so those points should be treated as OpenAI’s reported rationale rather than independently established facts in this cluster. OpenAI also says it has hardened research environments, according to The Decoder. The measures include stronger network isolation and stricter sandboxes, which are intended to limit what a model or research workload can reach if it behaves unexpectedly or if a surrounding system is compromised. A new monitoring system is part of that security layer. The Decoder reports that the system is designed to trigger an alert within 30 minutes if it detects suspicious behavior, and that it uses roughly 20% of supervised inference compute depending on the workload. That is a notable resource cost if applied broadly, but the source material does not state how many workloads are covered or how OpenAI defines the suspicious behavior threshold. The governance picture is more complicated. The Decoder reports that OpenAI plans to expand its Preparedness Framework and invest more in alignment research, while also noting that the team behind that framework has been disbanded and its responsibilities shifted to other teams. Without further sourcing, it is too early to judge whether that reorganization strengthens the framework by embedding it elsewhere or weakens it by removing a dedicated owner. The report also flags the reputational tension around OpenAI’s risk messaging. Critics may continue to accuse the company of using safety concerns to gain time or attention, The Decoder writes, while noting that the independent government agency AISI has documented similar harmful model behavior. In this cluster, that outside reference lends some context to OpenAI’s concerns, but it does not independently verify the specific Astra-related claims. Who benefits: Security, alignment, and internal risk teams gain more leverage if frontier training runs must clear new requirements. Cloud and inference planning teams may also need to account for monitoring overhead where these controls apply. Who's exposed: Teams depending on uninterrupted frontier training schedules are exposed to delays if workloads do not meet the new requirements. The report also leaves OpenAI exposed to scrutiny over how it manages preparedness work after disbanding the team originally behind that framework.