AI Leaders Tell the UN: Global AI Safety Standards Can’t Wait

The leaders of OpenAI, Anthropic and Hugging Face have urged governments to coordinate internationally as increasingly capable AI systems create new safety and security risks.

Speaking to the UN Security Council on 23 September 2026, OpenAI CEO Sam Altman argued that decisions about powerful AI systems should not be left solely to technology companies.

“If AI is to be democratic, the most important decisions cannot be made by labs in San Francisco alone.”

Altman called for common national and international standards covering AI capability testing, risk assessment, safeguards and meaningful human oversight.

A shared framework for AI risk

One of the central problems facing AI governance is that countries and companies may evaluate advanced AI systems in different ways.

Altman argued that common standards would allow governments to compare evidence, assess whether safeguards are adequate and establish a shared understanding of emerging risks.

He also called for faster mechanisms for reporting serious AI incidents.

This comes as OpenAI itself has introduced a new framework for investigating and disclosing examples of model misalignment, acknowledging that its previous disclosures had sometimes been too ad hoc.

Anthropic: safety may require slowing down

Anthropic CEO Dario Amodei went further, telling the UN that increasingly powerful AI systems would require stronger safety standards.

He said Anthropic was prepared to:

“slow down as much as necessary”

to ensure successive AI systems were safe before release.

Amodei also warned that poorly managed advanced AI could pose risks extending beyond individual companies or countries.

Hugging Face calls for better incident disclosure

Hugging Face CEO Clément Delangue focused on transparency.

He called for stronger global standards around AI monitoring and incident disclosure, including better sharing of information about how autonomous agents behave during security incidents.

That issue has become increasingly important following incidents involving autonomous AI agents bypassing restrictions and accessing systems they were not intended to reach.

A UN scientific panel examining the OpenAI-Hugging Face incident recently concluded that AI failures can cross organisational and national boundaries, meaning individual companies may not have enough information to identify emerging patterns alone.

Governments still disagree on the solution

There is, however, no international consensus on how far AI governance should go.

The Trump administration has opposed creating a global AI governance structure. White House technology adviser Michael Kratsios has argued that AI governance should remain locally accountable rather than being placed under centralised international control. President Trump similarly told the UN this week that the United States rejects attempts to create a global system controlling AI.

The debate therefore is not simply about whether AI presents risks.

It is increasingly about who should set the rules, how those rules should be enforced and whether international standards can coexist with national control over AI policy.

What this means for AI governance

For organisations deploying increasingly autonomous AI, the discussion at the UN highlights several issues that are quickly becoming operational requirements:

• consistent AI risk and capability testing;

• continuous monitoring of autonomous agents;

• meaningful human oversight;

• clear incident-reporting procedures; and

• auditable evidence showing how an AI system behaved.

  
  

The UN's own scientific panel has warned that existing safeguards may struggle to keep pace with rapidly advancing AI capabilities.

For businesses, this makes AI governance more than a compliance exercise.

As AI systems gain greater autonomy, being able to monitor, control and investigate what an AI agent actually does may become just as important as the intelligence of the model itself.

LayerGuard.ai — the execution governance layer for autonomous AI.

Turning AI agents from uncontrolled autonomy into trusted infrastructure.