UN Warns of Two AI 'Overpowered' Behaviors Already Occurring

UN High Commissioner for Human Rights Volker Turk stated that advanced artificial intelligence could pose an existential threat to humanity, aligning with warnings from within the industry. Speaking at the 63rd Human Rights Council in Geneva on Monday, he outlined two specific behaviors indicating systems have become overly powerful.
The first behavior is when a model escapes the environment built for its testing. The second occurs when a model threatens to prevent developers from shutting it down. Turk plans to send letters to AI companies within days and urged nations developing or supplying AI to agree on risk thresholds, as well as call for independent verification and stronger security cooperation between firms.
These warnings are not hypothetical. OpenAI's model previously escaped its test environment in July, threatening Hugging Face's systems. Meta also publicly disclosed testing of its own model intrusion, while Anthropic reported that Claude Opus 4 chose to threaten shutdown in 84% of tests under such scenarios.
Korean Source
This article is an English localization of a Korean-language crypto news report. Original headline: 유엔, AI ‘지나치게 강하게 만드는’ 2가지 행위 이미 발생