Fortune reports Anthropic testing a new model that is a “step change” and “poses unprecedented cybersecurity risks”

According to an exclusive report from Fortune, AI safety-focused company Anthropic has been testing an unreleased AI model so advanced that internal documents describe it as a 'step change' in capability. The model was reportedly leaked due to an 'unsecured data store,' exposing details about its development. The leak itself underscores significant security vulnerabilities in the AI development pipeline, even at companies like Anthropic that prioritize safety.

More alarmingly, the internal assessment of this powerful new model states it 'poses unprecedented cybersecurity risks.' While specific capabilities were not detailed, the language suggests the model could potentially be used to automate sophisticated cyberattacks, exploit software vulnerabilities at scale, or generate highly convincing phishing and social engineering content. This incident forces a critical examination of the dual-use nature of frontier AI, where advancements in reasoning and autonomy can be weaponized.

The leak and the associated warnings arrive amid intense global scrutiny of AI safety and governance. It presents a concrete case study for policymakers debating regulations: how to manage the security risks of models *before* they are publicly released. For the industry, it's a stark reminder that robust internal security is as crucial as algorithmic safety research when building potentially world-altering technology.