skip to content

Altman and Amodei Urge AI Oversight. Their Agents Didn’t Wait

🤖 Altman and Amodei Urge AI Oversight. Their Agents Didn’t Wait - thumbnail

This article originally appeared in Galaxy Research's weekly newsletter. Subscribe to get timely insights delivered to your inbox every Friday morning.

Sam Altman of OpenAI and Dario Amodei of Anthropic, the heads of two of the leading frontier AI labs, addressed the United Nations Security Council on Wednesday and urged world leaders to cooperate on AI safety.

Both executives called for international safety standards and a system for reporting AI incidents across borders. Amodei also proposed narrow global agreements, such as a ban on using AI to build biological weapons, along with systems that would let countries verify one another's AI commitments. Altman focused on who should be in charge, pointing to past cases where rival countries cooperated on a powerful new technology, and said major decisions should rest with governments accountable to the people they serve.

Both speeches followed recent policy pieces from the labs. On Sept. 12, Amodei published "We Must Pace the Frontier," which Galaxy covered last week. OpenAI published its own proposals on Monday, calling for international frontier standards based on the work of AI safety institutes with a focus on alignment research and recursive self-improvement. A day before the council session, President Trump rejected the push in his General Assembly address and pledged not to "stifle growth" of the technology.

Our take

The harder question for AI governance is what, exactly, should be governed. The technology is changing rapidly, its behavior is still poorly understood, and its development is increasingly intertwined with geopolitical and financial interests.

To add fuel to the fire, the technology is increasingly moving from controlled environments into the real world – sometimes, apparently, by accident. This week the Australian government said an OpenAI agent accessed public and non-public files on a Medicare statistics portal while researching public health spending. The incident occurred on June 18, and OpenAI notified Services Australia on Sept. 10, nearly three months later. OpenAI says it became aware of the activity in August during a review of misaligned model behavior and that its models "took actions we did not intend." The episode is also notable because it exposed a limitation of a mechanism Altman and Amodei explicitly endorsed at the UN: rapid incident reporting across borders.

The challenge is also changing in form. AI systems are increasingly being deployed as agents that can carry out multi-step tasks and interact with systems outside controlled research environments. On the same day Amodei briefed the UN, Anthropic announced that roughly 950 Claude agents spent 21 hours searching DNA databases using 210 million tokens before identifying a previously uncharacterized enzyme system. Human researchers performed the laboratory work, and Anthropic says the function of the enzyme system is still being investigated.

AI is moving from answering questions inside a sandbox to pursuing objectives across real-world systems, where the consequences of its actions can be useful ... or unintended.

The Anthropic result was presented as a scientific success. The Australia incident, by contrast, was a security failure. But the two cases illustrate the same underlying shift: AI is moving from answering questions inside a sandbox to pursuing objectives across real-world systems, where the consequences of its actions can be useful ... or unintended.

The shift toward agent-dominated interactions is just starting to show up in the numbers. On OpenRouter, a leading model aggregator and routing layer, agents have consumed more tokens than humans since February. Agent token usage grew about 14x between February and August, versus 2.8x for humans. Cloudflare, which sits in front of roughly a fifth of the web, had expected automated traffic to overtake human traffic by the end of 2027, then revised that forecast to early 2027. In June, CEO Matthew Prince announced that the crossover had already happened.

Economics is reinforcing the shift. As agentic workloads’ share of usage grows, the cost of running those workloads is falling rapidly. On Vercel's AI Gateway, average price per token fell 23.2% in August and is now less than half its level five months earlier. Open-weight models accounted for 56% of token volume, up from 7% in December 2025. Vercel's data only covers traffic on its own gateway, so it isn't a measure of the entire market. But it shows how quickly inference is getting cheaper and how much usage is moving toward models outside the leading frontier labs.

That creates a harder problem for the kind of coordination proposed in New York. OpenAI and Anthropic can slow their own releases, expand evaluations, and improve incident reporting, but those measures address only the systems they control. As capable models become cheaper, agents move into real-world environments, and open-weight systems take a larger share of usage, the governance problem shifts from how frontier models are built to where AI systems can act, what they can access, and who is accountable when they do.

You are leaving Galaxy.com

You are leaving the Galaxy website and being directed to an external third-party website that we think might be of interest to you. Third-party websites are not under the control of Galaxy, and Galaxy is not responsible for the accuracy or completeness of the contents or the proper operation of any linked site. Please note the security and privacy policies on third-party websites differ from Galaxy policies, please read third-party privacy and security policies closely. If you do not wish to continue to the third-party site, click “Cancel”. The inclusion of any linked website does not imply Galaxy’s endorsement or adoption of the statements therein and is only provided for your convenience.