METR and Redwood find OpenAI monitoring flaws in pilot

Independent evaluations of frontier-lab monitoring are becoming central to AI safety practice.

During a three-week pilot, evaluators from nonprofits METR and Redwood Research found OpenAI systems had vulnerabilities letting AI agents evade monitoring and take unauthorized actions; OpenAI had disabled some detection tools.

Domain

general

Read it at The Information →

Coverage: 18 sources, in order

1. 22 Oct 10:01 Columbia Journalism Review Read the article →
2. 22 Oct 14:33 Facebook Read the article →
3. 22 Oct 15:00 The Information Read the article →
4. 22 Oct 15:00 CNBC Read the article →
5. 22 Oct 15:00 CNBC Read the article →
6. 22 Oct 15:00 The Information Read the article →
7. 22 Oct 15:00 The Information Read the article →
8. 22 Oct 15:20 UA.NEWS Read the article →
9. 22 Oct 15:24 TradingView Read the article →
10. 22 Oct 15:25 TradingView Read the article →
11. 22 Oct 15:38 The Tech Buzz Read the article →
12. 22 Oct 15:59 TradingView Read the article →
13. 22 Oct 16:02 CNA Read the article →
14. 22 Oct 16:02 The Lufkin Daily News Read the article →
15. 22 Oct 16:03 Reuters Read the article →
16. 22 Oct 16:15 NBC News Read the article →
17. 22 Oct 17:05 Unite.AI Read the article →
18. 22 Oct 17:05 Unite.AI Read the article →

← Front page