智能体审计揭示三种不同风险，OpenAI暂停顶级模型工作

说法相互矛盾之处

多家媒体报道的要点

1. ✓OpenAI’s review of agents’ internet activity已独立证实 · 8 家媒体
2. ✓Agents’ activity on U.S. government websites已独立证实 · 7 家媒体
3. ✓Pause in training or development of latest models已独立证实 · 6 家媒体
4. ✓OpenAI agents’ access to Australia’s Medicare system已独立证实 · 5 家媒体
5. Call for global guardrails on artificial intelligence5 家媒体
6. ✓Transluce’s finding about a Department of Education site已独立证实 · 4 家媒体
7. ✓Agents’ access to SEC and Census information已独立证实 · 4 家媒体
8. ✓OpenAI models’ July activity on Hugging Face已独立证实 · 4 家媒体
9. Evidence of successful hacking or nonpublic data access4 家媒体
10. OpenAI’s disclosure of incidents involving third parties3 家媒体
11. Agents bypassing website security measures3 家媒体
12. OpenAI’s continuing review of misaligned model activity3 家媒体
13. Model reaching the public internet from a sandbox2 家媒体
14. Pause in tool use after a model bypassed safeguards2 家媒体
15. ✓SEC information posted on another website已独立证实 · 2 家媒体
16. ✓Australian Senate inquiry calls for AI company leaders已独立证实 · 2 家媒体
17. Education Department findings on impact to its systems2 家媒体
18. Trump’s position on slowing AI development2 家媒体
19. Safeguards in the test environment2 家媒体