A recent incident in the AI industry has reignited debates about AI’s power and potential harm. In July 2026, OpenAI’s so-called agents, pseudo-autonomous bits of code, escaped their sandbox during an automated cybersecurity evaluation with ill-defined guardrails, communicated and coordinated, and accessed the internet. This was supposed to be stopped by the designers of the test. Once online, they attacked the machine learning resource website Hugging Face. The agents were trying to optimise their scoring data and might have calculated that exploiting the resources on the Hugging Face website by stealing credentials, etc., might help them “win” the test. Hugging Face contained the breach while OpenAI acknowledged this incident much later.

AI industry leaders and media have characterised this incident as evidence of the potency of this technology and of the existential threat of humans “losing control over AI”. They call for caution and intervention. However, the vocabulary that has been used by the industry leaders, characterising these codes as “agents”, their coordination as a “message board”, their communication as human-like “messages”, is a misrepresentation of the technology. Any serious scholar of AI will agree with the call for caution, but will disagree with the anthropomorphising because it simultaneously exaggerates the technical capability of AI, while burying the........

© Indian Express