OpenAI tightens the screws on security to keep away prying eyes

OpenAI has significantly revamped its security protocols to safeguard against potential corporate espionage. The Financial Times reports that this initiative was accelerated following the emergence of a competing model from the Chinese startup DeepSeek in January. OpenAI has accused DeepSeek of improperly replicating its models through techniques known as 'distillation.' The updated security measures include stringent 'information tenting' policies that restrict employee access to sensitive algorithms and new developments. For instance, during the development of OpenAI's o1 model, discussions in shared office spaces were limited to only those team members who were officially authorized to engage in the project, as noted by the Financial Times. Additionally, OpenAI has taken further steps to protect its proprietary technology by isolating it within offline computer systems. The company has also introduced biometric access controls, including fingerprint scanning for entry into specific office areas. An internet policy that operates on a 'deny-by-default' basis has been established, meaning that any external connections require explicit approval. Moreover, OpenAI has enhanced physical security measures at its data centers and expanded its cybersecurity workforce. These changes not only reflect the company’s heightened concerns about foreign adversaries attempting to steal its intellectual property but may also address internal security challenges, especially in light of the ongoing competition among American AI firms and the frequent leaks of remarks made by CEO Sam Altman. OpenAI has been contacted for further comments on these developments.