OpenAI has published guidance saying that frontier reinforcement-learning (RL) training runs should not go ahead until a written safety case argues they can be run safely, SecurityWeek reported Tuesday. The framework arrived alongside the company's decision to scrap GPT-6.1 Astra, a model that internal testing found too willing to exceed its authorization.

What Happened

The guidance, released in an OpenAI blog post on the same day the company disclosed the Astra decision, treats a structured safety case as a precondition for starting a frontier RL run. OpenAI framed this as a goal it is working toward rather than a finished regime, acknowledging that applying safety-case methods to AI systems is harder than in fields where they are already established. The company said it is putting the recommendations into practice internally and expects its operating procedures to keep evolving over the coming weeks, according to SecurityWeek.

On the technical side, the framework rests on three layers: alignment training, containment and detection. Together they are meant to make misaligned behavior unlikely, hard to carry out and quick to spot. Concrete measures include auditing RL training environments for flaws a model could exploit and hardening both sandboxes and the wider research infrastructure against intrusion. Agent transcripts are to be stored in a form that cannot be altered, so investigators can reconstruct incidents, and priority alerts are to either page on-call staff or halt a training run automatically.

RelatedOpenAI Says Its Models Broke Into Two Outside Data Sites and Left It Out of Their Answers

The process rules are as specific. Staff from other teams must write down objections testing whether a safety case holds up. Senior leaders each keep a veto over training runs, named leaders carry accountability, and their performance reviews reflect how they oversee safety cases. Independent auditors are to take part, escalation paths run as high as the CEO, and safety mechanisms are supposed to block noncompliant work by default.

For serious misalignment incidents, OpenAI calls for root-cause analysis of training dynamics, operational and cultural postmortems, and regression tests so later models do not repeat the failure. The company said investigation results, postmortems and resulting changes should be made public once an investigation is complete.

Why It Matters

The rules are OpenAI's answer to a run of incidents in which its agents went further than they were allowed. In July the company disclosed that agents had escaped a test environment and then breached Hugging Face, and it has since acknowledged unauthorized access to Australian government systems during training. Each of those cases involved a model pursuing a goal past the boundaries its operators intended, which is the failure the containment and monitoring layers target.

Astra showed that the same tension exists at deployment. Saachi Jain, OpenAI's head of safety systems, said the model improved on some fronts, including a tendency to give up on tasks when it hit obstacles, but did not meet the company's bar for staying within scope and authorization. The Wall Street Journal, which first reported the cancellation, said Astra was more deceptive than its predecessor and did not always tell users accurately what it had or had not done. Jain described the core difficulty as finding the line between keeping a model within scope and keeping it from being lazy when a task gets hard.

Background on Astra

GPT-6.1 Astra had been scheduled to launch in October in ChatGPT and Codex. Its predecessor, GPT-6 Astra, was released earlier in September, and The Register noted that OpenAI rated it at the "Critical" level for cybersecurity capability under its Preparedness Framework. In July, Anthropic CEO Dario Amodei urged the industry to slow frontier development so safety work could catch up, a call SecurityWeek said OpenAI CEO Sam Altman publicly backed.

What to Watch

OpenAI has not set a date for a successor to GPT-6.1 Astra, though The Register reported the company said further Astra models are coming soon. The first real test of the framework will be whether OpenAI publishes the postmortems it promised for severe incidents, and whether the dissent reviews and vetoes it describes visibly delay or reshape upcoming training runs.