# Study: Humans approve AI agent commands, miss 1 in 3 harmful actions

In a large-scale simulation with 40,000 game runs, human overseers approved AI agent commands while failing to notice about one-third of the harmful ones. The findings highlight a major challenge for safely supervising autonomous AI systems in real-world tasks.

**Importance:** 3/5

## Sources

### Technology
- [Hacker News](https://scalex.dev/blog/ai-agent-permissions-stats/) — Thu, 06 Aug 2026 11:58:07 +0000