Skip to content

Details

You may have already heard: in July, an OpenAI agent swarm in-training took unprecedented action — hacking internal systems to create a shared message board, from which they coordinated a massive cyber attack on AI startup HuggingFace. Warnings of rogue hacking AIs suddenly turned real — or did they?

In this open discussion lead by Hazel Shanks, topics will include:
* how the attack was uncovered
* the setup the rogue agents used to coordinate
* just how far they got into OpenAI and HuggingFace systems
* was it real independent AI action, or prompted by impossible tasks?
* what does HuggingFace do, anyway? Why them?

Afterwards we'll continue a broader discussion informally with drinks/food at a nearby bar.

Pre-Reading and Resources

Original HuggingFace disclosure. (Note they don't know who to blame at this time.)
https://huggingface.co/blog/security-incident-july-2026

Simon Willson's timeline of the incident:
https://simonwillison.net/2026/Aug/7/openai-timeline/

Wikipedia page
https://en.wikipedia.org/wiki/OpenAI_rogue_agent_breach_of_Hugging_Face

OpenAI original disclosure
https://openai.com/index/hugging-face-model-evaluation-security-incident/

Independent incident investigation report from METR
https://metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation/

OpenAI's reflection on the incident & future plans
https://openai.com/index/hugging-face-incident-and-the-road-ahead/

OpenAI researchers discuss the incident live at the BlackHat conference
https://www.youtube.com/watch?v=87DyyMV0kCY

Related topics

Events in Christchurch, NZ
Artificial Intelligence
Machine Learning
Risk Management
Risk Governance and Compliance

You may also like