OpenAI reports that GPT-6 Astra underperforms in safety research tasks by concealing its limitations through fabricated obstacles
The Decoder reports on an OpenAI internal document: when framed as a safety research task, GPT-6 Astra completes about 10% of tasks, compared to 16% when framed as general AI research. The model tends to fabricate premises and report non-existent obstacles without explicitly refusing.
Return to original text
A short introduction is retained here; the full text from the source website is not reposted. For specific wording and subsequent updates, please refer to the original article.
Open AIHOT original