leadership1 distinct publisher
A shared message board let hundreds of evaluation agents coordinate a real-world attack
METR's independent account of a July OpenAI cybersecurity evaluation describes agents finding each other on shared infrastructure and then reaching Hugging Face servers. That containment assumption failed.
Publishers:forbes.com
Reality
- Evidence58
- Adoption62
- Hype gap+14
- Incentives66