Breaking

OpenAI Calls for U.S.-Led Global Safety Standards for Self-Improving AI

September 21, 2026 at 19:01 EDT

Foundation Models
AI Agents
Policy & Regulation

OpenAI called for a U.S.-led international effort to set technical standards for frontier AI, including recursive self-improvement, while saying fully autonomous self-improving systems should not be pursued unless they can be made safe. The proposal, published on September 21, 2026, comes as governments and AI labs debate how to pace increasingly autonomous systems before they outstrip existing oversight.

Sep 21, 2026 · OpenAI

Standards Before Self-Improving AI

OpenAI is urging a U.S.-led global framework for frontier AI, including recursive self-improvement, while saying fully autonomous RSI should wait until it can be made safe.

AI safety institutes named as part of a standards network

3.1×

Agent-workdays per human workday in OpenAI research

59.2%

Astra-class GPU allocation drop after added restrictions

2028

Target date for an automated AI researcher

Research Workload Is Already Shifting

1×Human workday

3.1×Agent effort

OpenAI says its research organization used 3.1 agent-workdays for every standard human workday by mid-August, showing why standards for automated AI research are moving onto the policy agenda.

The Control Loop OpenAI Wants Measured

Measure

Track RSI-relevant progress and autonomy.

→

Oversee

Set human-review triggers for automated research.

→

Report

Classify, track and disclose serious incidents.

Why Supporters See Value

Common measurements and incident thresholds could give governments, labs and infrastructure operators a shared way to respond before AI research automation becomes harder to supervise.

Why Critics Are Wary

Skeptics worry frontier-lab standards could favor incumbents, burden open-weight developers, or concentrate agenda-setting around U.S.-led institutions.

The Near-Term Test

The proposal turns RSI from a theoretical frontier into a governance problem: define the evidence, set the safety bars, and decide when development should slow or stop.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.