China's AI model raises fresh US security concerns over cyber vulnerabilities

Anthropic, a leading AI safety firm, has raised alarms about China's GLM-5.3 model, warning that it demonstrates capability to identify security flaws in systems and construct end-to-end cyber exploits. The concerns have thrust discussions about AI governance and international technology competition back into focus amid escalating US-China tensions over advanced technology development.

A key vulnerability flagged by researchers centres on the model's open-weight architecture, which allows outside developers to modify its built-in safety guardrails more readily than proprietary alternatives. This accessibility raises the prospect that the AI system could be adapted to bypass protections designed to prevent misuse for malicious cyber operations.

The disclosure highlights a broader policy dilemma confronting technology regulators worldwide: balancing the benefits of open-source AI development against the risks posed by unrestricted access to potentially dangerous capabilities. The incident underscores growing concerns that advanced AI systems could inadvertently—or deliberately—enable sophisticated cyberattacks if safeguards are circumvented.

The situation intensifies existing scrutiny on China's artificial intelligence advancements and their implications for global cybersecurity. US policymakers have increasingly scrutinised Beijing's technology sector as competition for AI dominance accelerates, with officials warning of potential security implications for critical infrastructure and defence systems across allied nations including India.