エージェント監査で3つの異なるリスクが浮上、OpenAIが最上位モデルの作業を一時停止

報道が食い違う点

複数のメディアが報じた論点

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