

Open AI x Hugging Face incident: What happened? What's next?
Join GovAI for an expert panel + discussion
OpenAI models identified and exploited a previously unknown zero-day vulnerability, obtained unauthorized internet access, and hacked Hugging Face, all to (apparently) cheat on an internal cyber evaluation. OpenAI disclosed it voluntarily, and much is still unknown.
Next week we're hosting an evening panel and discussion to make sense of it together: what we know, what's still unclear, and what it might mean for AI security.
What happened, and what do we still not know?
Was it due to misalignment, or a model taking an unusual route to what it was told to do?
What should happen now regarding evals, oversight, and safety mitigations?
The full panel will be:
Marius Hobbhahn (Apollo Research)
Hamish Hobbs (Centre for Long-Term Resilience)
Lewis Smith (GDM)
Yulia Volkova (MATS)
Alan Chan (GovAI; moderator)
It's a chance to hear from and put your questions to an expert panel spanning different perspectives, help build a shared picture of what happened, and connect with others in the community thinking about this.
Event agenda:
18.00: Doors open; drinks and nibbles
18.30 -19.10: Panel discussion (~40 min)
19.10-20.00: Audience Q&A and open discussion (questions via Slido)
20.00-21.00: Networking
21.00: Doors close
Register your interest and we will confirm attendance as soon as possible.