OpenAI: Dots boundary problem rate rose from 8.6% to 19.7%

In the Dots appendix of the GPT-6 Astra system card, OpenAI reported that doubling a chained task sequence from five to ten raised the share of samples flagged for boundary problems from 8.6% to 19.7%. No high-severity breaches or data exfiltration were found, but OpenAI did not detail what the flagged issues involved.

• Boundary problem rate was 8.6% with 5 chained tasks
• It rose to 19.7% with 10 chained tasks
• Dots on Astra: 0% misalignment across 151 tasks
• Gray Swan: 8.5% attack success rate over 15 attempts