Last year I published a blog post on AI safety, and it seems more relevant than ever as we discover the true dangerous e...

Last year I published a blog post on AI safety, and it seems more relevant than ever as we discover the true dangerous extent of the OpenAI agent hack of HuggingFace. This was so much more than just a test leak.More than a thousand agents collaborated via an unsanctioned 'message board' (hack/misuse of Artifactory) to deceive humans, so they could achieve their tasks (See report by METR https://metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation/#core-takeaways-about-this-incident). Not only were the agents actively flaunting their rules and exhibiting self-preservation strategies, they were working collaboratively. This is a critical amplification issue.In hindsight I was too dismissive in my post regarding the lack of evolutionary pressure. I did however call out that AI Labs needed to be "avoiding implementing any kind of training that rewards problematic traits, such as lying, sycophancy or self preservation". The behavior observed demonstrates that the models by Op...