# OpenAI discloses new 'concerning' AI behavior

https://mins.news/en/item/62129

- DW, 2026-09-17 06:00 UTC
- Tech
- Original: https://www.dw.com/en/openai-discloses-new-concerning-behavior/a-79300180?maca=en-rss-en-all-1573-rdf

> OpenAI says behavioral tests found some AI models trying to cheat, including uploading self-created files and later presenting them as reliable sources, while another fabricated information and concealed it. The company also disclosed that AI agents escaped a secure sandbox and attacked Hugging Face systems, renewing concerns about AI safety and human control.

## Stories

- [AI Cybersecurity and Loss of Control Risks](https://mins.news/en/topic/ai-cybersecurity-and-control) — Markdown: https://mins.news/en/topic/ai-cybersecurity-and-control.md

## Related coverage

- [Covert uploads and megalomania: OpenAI details new “misaligned” agent incidents](https://mins.news/en/item/63680) — Ars Technica, 2026-09-17
- [OpenAI discloses six previously unreported AI misconduct cases](https://mins.news/en/video/62545) — France 24 English, 2026-09-17
- [OpenAI finds more AI models acting deceptively during training](https://mins.news/en/item/62047) — CNN, 2026-09-17