The Guardian view on Anthropic’s Claude Mythos: when AI finds every flaw, who controls the internet? | Editorial

Editor's take

Anthropic's Claude 3 family of models, particularly Opus, has demonstrated a significant capacity to identify vulnerabilities in online systems. This development underscores the dual-use nature of advanced AI, where powerful tools for defense can just as readily be leveraged for offense. As these models become more accessible, the concentration of such potent capabilities within a few private companies like Anthropic raises concerns about who ultimately wields influence over the internet's security and integrity.

The implications extend beyond mere cybersecurity; this capability touches upon the very governance of the digital sphere. If private entities can develop AI that quickly exposes systemic weaknesses, it creates a power imbalance between those who control the AI and the broader public or even governments who must then react. This mirrors ongoing debates about AI's role in misinformation and censorship, suggesting a pattern of private AI development shaping public digital infrastructure.

Future developments to monitor include the emergence of open-source models that can match or surpass Claude 3's vulnerability discovery speed, and the regulatory responses from governments worldwide. The speed at which these capabilities are democratized or controlled will be crucial in determining whether AI reinforces existing power structures or fosters a more distributed approach to digital security and access.