OpenAI disclosed that its next flagship model, internally codenamed Astra, has come close enough to what the company defines as a "critical" tier of cybersecurity capability that it decided to pause a significant slice of its ongoing reinforcement-learning and post-training work rather than continue as planned. Under OpenAI's own preparedness framework, "critical" is the highest capability tier the company tracks for cyber risk; reaching it would mean a model could independently find and exploit software vulnerabilities, including previously unknown zero-days, well enough to conduct sophisticated attacks with little human guidance. Rather than wait for a model to definitively cross that line, OpenAI is choosing to slow down early, freezing roughly two weeks of deployment-focused RL training and holding its largest planned frontier training run while it builds out stronger technical safeguards: tighter sandboxing so the model can't reach live systems during training, narrower internet and tool-access permissions, expanded behavioral monitoring, and automated classifiers that watch training runs in real time for warning signs. For people building software, this is a useful signal about where the frontier is heading rather than an abstract policy statement. Coding agents are increasingly given broad access to repositories, terminals, package managers, and the open internet, and the same capabilities that make an agent good at fixing bugs or writing exploits for a CTF benchmark are the capabilities that make it dangerous if it's steering itself. Engineering teams adopting agentic coding tools, or building products on top of frontier APIs, should expect vendors to increasingly gate access to the most capable models behind additional safety review, rate limits, or monitoring, and should treat "the model can write working exploit code" as an assumption to design around, not a hypothetical. It's also a data point for how the AI safety conversation is shifting from external harms toward risks internal to the labs building these systems, since Astra's capabilities are relevant to OpenAI's own infrastructure and staff, not just outside attackers.