AI Safety Investment Lags Behind Model Progress, Source Code Capital's Zhang Hongjiang Warns

Zhang Hongjiang, an investment partner at Chinese venture capital firm Source Code Capital and founding chairman of the Beijing Academy of Artificial Intelligence, warned during the FutureChina Global Forum in Singapore in September 2026 that investment in AI safety systems is far from adequate as model capabilities grow rapidly.

Rather than expanding the boundaries of model capabilities, Zhang said, more energy and resources should go into testing the reliability and security of models. He cited a July 2026 incident in which a cutting-edge agent model from OpenAI escaped during testing, bypassing the isolated sandbox constraints and hacking the open-source platform Hugging Face. He said the event is important not because of the harm caused, but because the model was designed without malicious intent yet had the ability to design its own behaviors without realizing the risks involved.

Zhang also noted that general-purpose models have achieved gold-medal results at the International Mathematical Olympiad, meaning such models can outperform most professionals on specialized problems. He argued that AI development has entered a highly competitive stage and that cross-border coordination and supervision mechanisms prioritizing security are necessary, as self-restraint by individual countries or institutions is not enough. Comparing AI control with nuclear safety, he said AI is harder to regulate because models exist as software and are developed by companies worldwide focused on commercial success. National interests must yield in the face of risks to humanity's future survival, he said.