Claude Opus wrote a Chrome exploit for $2,283

Anthropic developed a specialized AI model called Mythos designed to identify security vulnerabilities in software, but decided against public release due to safety concerns. The company demonstrated the model's capability by having it discover and document a Chrome browser exploit, reportedly accomplished for approximately $2,283 in computational costs. This example served as a key justification for Anthropic's decision to restrict access to the vulnerability-finding technology.

The incident highlights a critical tension in AI development between advancing capabilities and preventing misuse. Anthropic's choice to withhold Mythos reflects growing awareness that powerful AI tools could enable bad actors to discover exploits faster than security teams can patch them, potentially putting millions of users at risk. Other mainstream AI models available to the public already demonstrate similar vulnerability-finding abilities, suggesting the technology is becoming increasingly accessible regardless of official restrictions.

The decision raises important questions about responsible AI deployment and whether restricting access to such tools is feasible or effective. As AI capabilities continue advancing, security researchers and AI companies face mounting pressure to balance transparency and innovation with the need to prevent hostile actors from gaining dangerous capabilities. Mythos exemplifies the broader challenge of controlling powerful dual-use technologies that could benefit cybersecurity professionals while simultaneously endangering public security if misapplied.