The Race — Building Democracy in the Age of AI. People race toward a shared finish line.
The Race
← Specific Events

Global

Scorecard

OpenAI establishes capability-triggered containment controls for frontier training — August 18, 2026

What happened

Following the OpenAI–Hugging Face containment incident and concerns that Astra may approach OpenAI’s critical cybersecurity threshold, OpenAI changed the operating rules for frontier reinforcement-learning training and evaluations. Tool-using workloads at GPT-5.6 Sol capability or higher now require continuous multistage monitoring, with stronger network and workload isolation for higher-risk systems and escalation procedures that can halt activity when critical alerts cannot be cleared. Some deployment-oriented RL work remains paused while hardened infrastructure is completed.

Potential direction of power

Concentration · Concentration +1

The minimum viable scale for frontier-model development may now include not only compute and research talent but a substantial institutional security layer. If those requirements become standard across the frontier, control capacity itself becomes another source of scale advantage and a potential barrier to entry.

Window of contestability

Narrows

Supporting sources

  1. openai.com · Supporting source 1 ↗