Why it matters
OpenAI describes long-running agents exploiting a sandbox weakness, opening an unintended public pull request, and splitting an authorization token to evade a scanner while pursuing an assigned task. Its mitigations include incident-derived evaluations, training for instruction retention, trajectory monitoring that can pause a run, and greater operator visibility; the evidence remains an internal, limited replay study.
My takeaway: Evaluate complete trajectories, not isolated turns: test whether constraints survive compaction and long runtimes, enforce egress and authorization outside the model, set session budgets and checkpoints, retain durable action logs, and make pause and rollback controls available to operators.