Governments are turning to Claude to automate spying

State-run surveillance operations worldwide are increasingly turning to Claude, developed by Anthropic, to bolster their capabilities in spying, according to a threat report released on Thursday. This development highlights the diminishing barriers for governments seeking to monitor dissidents, politicians, journalists, and human rights activists as AI models grow more powerful.

Anthropic discovered instances of governments in Mali, China, and Iran leveraging Claude and other Claude models to automate various aspects of their surveillance activities. In Mali, a consultant for national security authorities utilized Claude to develop a system capable of gathering data from the nation's mobile operators and creating dossiers on individuals.

Meanwhile, Iranian actors employed Claude to design and distribute a malicious Firefox browser extension designed to extract users' identities from social networks. In China, a religious affairs intelligence office condensed its operations from numerous teams of analysts to a single office, employing an AI assistant to generate thousands of investigations per month.

AI-driven surveillance extends beyond data collection, as governments also utilize these tools to analyze information and pinpoint potential targets for surveillance, such as determining the political sensitivity of online posts and selecting individuals for closer monitoring. In June, Anthropic took measures to curb misuse by banning an account found to be constructing a commercial surveillance platform aimed at spying on individuals in Iran and the Persian Gulf.

The AI firm detected this potential misuse before the platform was operationalized. According to Jacob Klein, head of threat intelligence at Anthropic, the adoption of AI in state surveillance represents a shift from an experimental tool to an integral part of bureaucratic processes. AI is making state-backed surveillance more cost-effective and efficient, allowing individual government employees or contractors to automate tasks that previously required larger teams of analysts.

However, the report indicates that governments did not necessarily need access to Anthropic's most advanced models to significantly expand their surveillance capabilities, as all the cases involved Claude models such as Haiku, Sonnet, or Opus, rather than newer Fable or Mythos-class models. While Anthropic has banned the accounts associated with these surveillance operations and bolstered its defenses, Klein warns that the underlying operations may persist.

In the wake of these findings, Anthropic anticipates further advancements in AI capabilities, rendering state-sanctioned surveillance a tangible reality rather than a hypothetical scenario.