AI Safety Approaches Diverge as Anthropic Expands Security Testing

The expansion of Anthropic's 'Project Glasswing' to more security teams, which offers access to its powerful AI models with reduced safeguards, signals a distinct approach to AI safety. The initiative, active since April, aims to identify software vulnerabilities. Between April and July, its partners found 129,000 verified vulnerabilities. Anthropic's internal open-source scanning efforts located an additional 5,500 vulnerabilities from April to October. These findings underscore the ongoing challenges in securing software systems, even as AI models become more integrated.

This strategy, prioritizing broad exposure for vulnerability detection, contrasts with other industry discussions around AI model access. Some U.S. Big Tech companies advocate for American open-weight models, while Anthropic's CEO has expressed disagreement regarding cybersecurity risks associated with such openness. This divergence highlights a fundamental tension: the push for accelerated AI development versus the imperative of maintaining digital security. China, for instance, has demonstrated a lead in open-weight models, according to CNBC in July, suggesting varying national approaches to AI development and deployment.

The deployment of advanced AI, even for security purposes, introduces complex dynamics into national and regional digital infrastructures. For nations across Asia, including Thailand, the proliferation of powerful AI tools, whether open or closed, necessitates a re-evaluation of cybersecurity postures. The findings from Glasswing — over 134,000 vulnerabilities identified by October — indicate the scale of the task. If AI itself is used to uncover these flaws, it also points to the potential for AI-driven threats. The New York Times, in August, observed that AI is becoming so powerful that it stumbles those attempting to contain it.

The broader implication for the Asia-Pacific region lies in how different economies choose to regulate or engage with such AI initiatives. Will the emphasis be on fostering open AI ecosystems, potentially accelerating innovation but also exposing systems to greater risks? Or will a more restrictive approach prevail, prioritizing control and stability? The approach Anthropic has taken with Glasswing—allowing vetted professionals to test models with fewer safeguards—is a specific example of balancing these competing demands. Its success or failure in mitigating vulnerabilities will inform future policy considerations across a region increasingly reliant on digital infrastructure. The unresolved tension remains how to balance the benefits of rapid AI deployment with the escalating complexity of cybersecurity in diverse technological landscapes.