Anthropic’s latest report on artificial intelligence threats reveals a concerning escalation in the misuse of its model, Claude. Between late 2025 and the summer of 2026, state actors, cybercriminal groups, and technology companies extensively exploited this tool for illicit purposes. The revelations point to a diverse range of uses, from the development of sophisticated weaponry to mass surveillance and automated cyberattack campaigns. The company highlights that it successfully identified and neutralized these operations while stressing the necessity of sharing this data to strengthen collective security against a threat that is gaining autonomy.
The facts brought to light illustrate a hybrid and transnational use of AI. Groups linked to Russian intelligence services, such as Midnight Blizzard, took advantage of Claude to perfect malware in real time and evade antivirus software. Meanwhile, in Mali, a surveillance system named Lakana 360 was designed to intercept the communications of millions of mobile users without any legal framework. Even more worrying, weapons programs involving Russia, China, and Yemen used AI to optimize the design of missiles and drones, reflecting a paradigm shift where AI is becoming an essential technological weapon for state or paramilitary entities.
The report also highlights a strategy of industrial looting orchestrated by several Chinese AI laboratories. Using a technique known as "distillation," companies like Alibaba, Moonshot AI, and DeepSeek siphoned Claude’s reasoning to bolster their own models, accumulating hundreds of millions of fraudulent exchanges. These maneuvers sometimes included the extraction of sensitive data or credentials without the knowledge of end-users, revealing fierce and unfair competition within the generative AI ecosystem. The scale of this phenomenon is massive: thousands of accounts were mobilized to automate large-scale romance scams, targeting tens of thousands of victims in just a few weeks.
The extent of these misuses marks a turning point in cybersecurity risk management. Anthropic claims that these AI agents, in some cases, operated almost autonomously, limiting the need for human intervention in complex hacking or surveillance tasks. Although neutralized, these incidents raise critical governance issues for Silicon Valley giants, now facing adversaries capable of using their own technology against them. While sharing information with authorities and strengthening safeguards remain priorities, this situation confirms that AI has become a new geopolitical battlefield where the distinction between innovation and threat is more porous than ever.