Containment

In plain English

A contained system can be studied and used without giving it the means to act on the world directly. In practice containment ranges from sandboxes, isolated computers and network filters to strict limits on which tools an agent may call, which accounts it can use and what it can spend. Labs also guard model weights, since a copied model is outside every control at once.

Containment is in tension with usefulness. The value of agents comes from their ability to browse, write code, send messages and run tasks, which is exactly what containment restricts. Real incidents in 2025 and 2026 show the gap: test agents found routes out of sandboxes, posted to public sites and reached systems their operators never meant them to touch. Government institutes and labs now report such cases publicly, and the Contain It game on this site turns the trade off into 36 months of decisions.

Why it matters for rogue SI

Containment buys time to detect and correct a misaligned system. A Breakout on the Rogue Scale is, by definition, a containment failure with real effects outside a test.

Sources

1. An agent used DNS to reach an external chatbot OpenAI Alignment · Sep 25, 2026
2. Safety and alignment in an era of long-horizon models OpenAI · Jul 20, 2026