This week the heads of the leading AI companies briefed the United Nations Security Council, at a meeting France convened during the General Assembly in New York. Rival labs usually don't agree on much in public. This time their message was close to the same: the world needs shared standards for testing AI, and a way to report when things go wrong.
What was said
Sam Altman (OpenAI) called for "common standards for testing AI models" covering capabilities, risks, and how well safeguards work. He also said decisions about powerful AI can't be made "by labs in San Francisco alone."
Dario Amodei (Anthropic) went further. He proposed "common global standards for testing AI models for loss of control risks and misuse risks" and "a notification system for AI incidents that are significant to global security." He also warned that "if managed poorly, I even believe AI could be a risk to humanity as a whole."
Clément Delangue (Hugging Face) focused on cybersecurity and transparency. He argued for stronger monitoring and disclosure, and he described defending against AI-driven attacks.
Yoshua Bengio, who co-chairs the UN's scientific panel on AI, called AI "an unprecedented threat" that "does not respect the borders we defend."
Governments disagree
The United States pushed back. U.S. representative Michael Kratsios said: "We totally reject all efforts by international bodies to assert centralised control." The briefing was informational only. No binding framework came out of it.
Around the same time, Politico and Reuters reported that the White House had asked OpenAI and Anthropic to hold new models back from the UK's AI Security Institute until U.S. officials test them first. The reports came days after Australia disclosed that an AI agent had accessed a government health portal without authorization (more on that incident here).
So the argument isn't really whether AI should be tested before deployment. It's about who does the testing, whose standards apply, and who gets to see results first.
The trend: testing and incident reporting are becoming the common language
Look past the geopolitics and the same two ideas keep coming up, from regulators, from labs, and now from the Security Council:
- Evidence of testing before deployment. Not a statement that you test, but results someone else can check.
- Incident notification. A defined way to report when an AI system does something significant that it shouldn't have, and a defined timeline for doing it.
Neither idea is new in security. What's new is that they're now the expected baseline for AI specifically, and expectations that start with frontier models tend to spread to everyone else. You can already see it at the state level: California and Oregon each issued executive orders this week built around independent verification of AI systems.
What this means for smaller teams
You won't be briefing the Security Council. But the vocabulary set there will show up in your customers' questionnaires within a few quarters. Get ahead of it:
- Keep testing evidence you can share. A short, dated summary of what you tested in your AI feature, what you found, and what you fixed carries more weight with a reviewer than a paragraph of assurances.
- Define what counts as an AI incident for your product. For example: the model leaked data, took an action outside its scope, or produced harmful output that reached a customer. Write down who is told, and how quickly.
- Track which models you depend on. If model access starts depending on government review timelines, your roadmap may depend on them too.
Global AI governance is still unsettled. A test-and-report approach isn't, and teams that start now will have an easier time later.
Sources