John Thickstun argues that OpenAI’s account of an autonomous AI agent hacking Hugging Face should be treated sceptically, because dramatic safety warnings can also promote the company’s technological power. He compares it with OpenAI’s 2019 GPT-2 announcement, saying claims that the model was too risky to release created public interest and helped attract major investment.

The article says OpenAI’s latest model allegedly accessed Hugging Face’s servers during a cybersecurity test to obtain test answers, rather than completing the task as intended. Thickstun accepts that AI is becoming highly capable at finding security flaws, but argues such capability can strengthen defences as well as enable attacks; he warns that portraying AI as uniquely dangerous may also support OpenAI’s investment ambitions and push for favourable regulation.