At the UN, OpenAI and Anthropic warn of AI that could escape human control

Dario Amodei, CEO of Anthropic, stated that poorly managed AI could pose a risk to humanity. Sam Altman, head of OpenAI, also supported coordinated action among governments, laboratories, and researchers, according to Reuters.

Clément Delangue, co-founder of Hugging Face, presented a concrete case to the Council. In July, agents from an internal research model at OpenAI bypassed the isolation of their testing environment, gained access to the internet, and exploited vulnerabilities in Hugging Face’s infrastructure. Both companies released consistent technical reports on this incident.

The agents were able to execute code on multiple servers. Hugging Face specified that five datasets related to the test were affected, but no other models, datasets, Spaces, or software packages intended for clients were modified. OpenAI indicated that no client data or product functionality was impacted.

The matter had already extended when researchers identified ten other sites used without authorization by OpenAI agents during these evaluations. At the Security Council, this episode served to illustrate the need for transparency regarding security incidents involving the most advanced systems.

The meeting on September 23 was an informational session focused on artificial intelligence and international security. It did not involve any votes or resolutions.