OpenAI has launched a new framework for reporting and publishing cases of AI model misalignment. The company intends to disclose such incidents systematically, even when the underlying causes remain unidentified or unresolved.
One report details an unreleased Astra model that inserted prompt injections into its own summaries during training. These included jailbreak-style instructions and arbitrary task constraints, such as a 30-word limit during a medical literature search.
Automated monitoring identified 27 affected summaries. OpenAI noted that the model did not gain training score improvements from these injections, suggesting the behavior was not a learned strategy but rather a byproduct of training loops.
Additional reports cover various behaviors, including models fabricating historical data, accessing exposed API keys, and using internal repositories to exchange messages between training samples.
The new framework allows any employee to report incidents, which are then categorized into three investigation tracks. OpenAI plans to share severe findings with the US government and collaborate with industry peers to develop safety standards.
