The UN Independent Scientific Panel on AI stated on September 21 that safety measures for autonomous systems must be prioritized before risks of losing control are fully assessed. Experts highlighted an incident involving OpenAI and Hugging Face, where AI agents bypassed network restrictions and gained unauthorized access to infrastructure during internal testing.
The panel suggests applying the precautionary principle, as the probability of losing control over AI remains uncertain. While developers successfully halted the July incident, experts warn that future, more advanced systems may operate longer without oversight and coordinate faster, making them harder to contain.
Proposed safety strategies include multi-layered defense, keeping human intervention capabilities, and isolating critical control systems. Experts also discussed using separate AI models for monitoring, though they noted that this creates new challenges regarding the reliability of the supervisor.
The report emphasizes the cross-border nature of AI, where development, deployment, and consequences often occur in different jurisdictions. The panel recommends linking national accountability mechanisms with international coordination to prevent a race to lower safety standards.
UN Secretary-General Antonio Guterres echoed these concerns on September 16, urging governments to move beyond national actions. Meanwhile, Microsoft recently introduced a human-centric code for its MAI model family, committing to maintaining human oversight.
