OpenAI and Anthropic have published policies that tie deployment of some high capability AI systems to safeguards against severe harm.

OpenAI's Preparedness Framework says covered systems that reach its High capability threshold must have safeguards that sufficiently minimize associated severe risks before deployment. Anthropic publishes a Responsible Scaling Policy with capability thresholds, required safeguards, and risk reporting. These commitments are evidence that major labs recognize the issue and have governance processes intended to address it.

Whether those processes will deliver effective protection before agents acquire catastrophic capabilities is a separate forecast. That forecast needs evidence about implementation, adversarial performance, scope, and whether commitments remain in force under real deployment pressure. The July 2026 OpenAI incident gives a concrete reason to ask how safety policies perform when agents exploit unexpected routes.