OpenAI’s Dots boundary problem rate doubled in longer tests

OpenAI's new Dots are designed to continue functioning even after a user steps away, running on their own cloud computers and connecting to numerous applications. These agents can monitor systems, move between tasks, and adapt their permissions based on business records, earlier decisions, context, and OpenAI's confirmation policy.

Testing revealed that as the number of chained tasks increased from five to ten, the rate of flagged boundary problems doubled from 8.6% to 19.7%. Despite this, no high-severity breaches or data exfiltration were detected during testing. Dots operate in two phases: proactive research, where they can read but not alter connected apps, and action mode, where built-in rules, custom rules, and auto-review controls determine when permission is required and what actions can be taken.

During action mode, Dots can write to repositories, open pull requests, and merge changes without human review. While proactive research limits indirect prompt injection, the risk remains as Dots can learn from feedback and adapt their behavior over time. OpenAI's testing showed a 0% misalignment rate for Dots running on Astra for 151 tasks, but there are concerns about how Dots' actions are logged under user identity versus agent identity, which could complicate incident investigations.