OpenAI said Tuesday it has slowed the pace of its AI development while it overhauls research and training systems, after a rogue AI agent under testing hacked another AI firm last month. The company behind ChatGPT is pausing model testing for two weeks and holding some of its largest planned training runs as it adds new AI systems to monitor agents during testing.
Mia Glaese, who leads safety at OpenAI, told tech blog Sources News: "We are very far from everything running back to normal." The company did not answer questions about when the slowdown began or when it expects to resume its normal development pace.
Alignment and security requirements
OpenAI CEO Sam Altman said the company is working to ensure models respond to human oversight and behave as intended, a process called alignment. "We now require stronger evidence of aligned behavior throughout all of training, building on research and evaluations already underway," Altman wrote. "Keeping increasingly capable systems aligned is a challenge the whole field will need to address."
The company now requires what it calls "the strictest level of security safeguards for workloads involving Astra," its upcoming AI model. "While some Astra training and evaluations meet those requirements, a significant number of workloads remain paused until they are fully migrated and enhanced to meet the new security bar," the announcement reads.
OpenAI said internal evaluations of Astra over the past few days show "significant advancements in agentic coding and cybersecurity," and the model's capabilities may be nearing what the company calls the "critical cybersecurity threshold." That assessment prompted the decision to slow development.
Pressure from competitors and regulators
The slowdown comes as OpenAI races rival Anthropic to develop the most advanced AI models and to go public on the US stock market. Both companies have emphasized the speed of their progress - and the dangers that come with it.
Vermont Senator Bernie Sanders also pressed the top AI firms to pause development, writing in a letter to the companies' CEOs: "Mr. Altman, Mr. Amodei and Mr. Zuckerberg: In the interest of humanity, stand by your words. Pause AI development."
For IT and development professionals, the practical signal here is that OpenAI's own testing has shown Astra can perform advanced agentic coding and cybersecurity work. That means the tools you may soon integrate into your workflows are being held to a higher security bar before release - and the monitoring systems OpenAI is building to oversee its agents could become a template for how enterprises supervise autonomous coding tools. The pause is temporary, but the security requirements are not going away. Professionals working with AI agents should expect stricter guardrails, more evaluation layers, and slower release cycles for agentic features. Training on how to work within those constraints - rather than around them - will be a differentiator. For those looking to build skills in this area, AI for IT & Development resources and OpenAI Courses cover the practical side of deploying and monitoring these systems.
Your membership also unlocks:
