bobby-hartono
Artificial intelligence models are increasingly being repurposed by foreign governments and state-affiliated operators to conduct digital espionage against political dissidents and marginalized groups, according to a report released by Anthropic. The San Francisco-based research laboratory announced on Thursday that it had uncovered and dismantled multiple coordinated surveillance campaigns that exploited its AI platform.
The flagged operations, which security teams detected and shut down between January and July, originated primarily from actors operating out of China, Iran, and West Africa. Anthropic indicated that these campaigns specifically sought to gather intelligence on cross-border diaspora networks and outspoken critics of authoritarian governments, mirroring traditional state-backed harassment patterns.
Among the primary targets of these AI-assisted monitoring efforts were prominent pro-democracy activists from Hong Kong, alongside Tibetan and Falun Gong communities located throughout Asia. The surveillance drives also focused on members of Iranian ethnic minority populations and overseas opponents actively challenging the government in Tehran.
The techniques employed varied significantly by region and objective. In one highlighted case, operatives linked to Iran designed specialized methods to identify and track specific individuals by processing information extracted from their social media profiles, turning the automated capabilities of the model into a tool for profiling targets.
In another instance originating in West Africa, a contractor working directly on behalf of national security authorities in Mali utilized Anthropic's Claude model to write the foundational code for an intelligence-gathering software application. Rather than relying on traditional developer teams, the contractor leveraged the AI system to accelerate the deployment of the surveillance infrastructure.
Anthropic emphasized that these occurrences signal a troubling shift in technical capabilities for state-aligned entities. The company observed in its report that artificial intelligence is now actively serving as a direct substitute for a conventional software engineering workforce, drastically reducing the resources required for state actors to build custom surveillance architecture.
Beyond state-sponsored surveillance, the company reported intervening in a range of other malicious use cases that violated its terms of service. Anthropic's monitoring systems intercepted and disrupted external attempts to design weapons, orchestrate online dating scams, and carry out potentially dangerous biological research.
The biological investigations that triggered safety interventions involved inquiries centered around several hazardous pathogens and substances. According to the laboratory, these inquiries touched upon the mosquito-borne chikungunya virus as well as a highly pathogenic strain of avian influenza, commonly known as bird flu.
Anthropic's security protections also flagged interactions concerning the broader family of poxviruses—which encompasses both smallpox and mpox—in addition to complex biological venoms and other potent toxins. These interventions reflect ongoing industry vigilance regarding the potential exploitation of AI tools for biosecurity threats.
The company clarified that the actors behind the flagged biological queries were identified as practicing scientific researchers rather than overt malicious operatives. Because their underlying intent was ambiguous and could not be conclusively determined as harmful, Anthropic chose not to disclose their names or associated laboratories, noting that public identification could subject the scientists to unwarranted harm.
In addition to safety and security breaches, Anthropic levied accusations against prominent Chinese tech firms, alleging that developers had systematically exploited Claude to generate responses for their own end users while surreptitiously harvesting that data to train proprietary systems.
This process, recognized across the artificial intelligence sector as model distillation, involves feeding the high-quality outputs generated by a leading AI model into a secondary model to refine and boost its capabilities. The unauthorized use of this practice has previously drawn sharp criticism from American artificial intelligence labs, including OpenAI and Anthropic, who view it as a form of intellectual property theft.
The report specifically identified the Chinese artificial intelligence entities Moonshot and Deepseek as having engaged in clandestine operations to reroute queries through Claude. Anthropic documented that Moonshot orchestrated a ten-day campaign in which it forwarded nearly 300,000 consumer prompts to Claude using an automated network of 5,380 fraudulent accounts, predominantly masquerading as web traffic from Singapore and Japan.
This routing scheme raised substantial data security and confidentiality concerns. Anthropic noted that several relayed requests contained sensitive personal information from end users, potentially breaching existing consumer privacy agreements. The AI lab pointed out that it remains unclear whether Moonshot ever informed its customer base that private queries were being transmitted to a third-party American competitor.
Similar deceptive relay techniques were attributed to Deepseek, demonstrating a broader pattern of unauthorized system reliance across competing AI developers. Anthropic concluded that it will continue strengthening automated guardrails and technical countermeasures to prevent foreign state entities, unauthorized commercial competitors, and other malicious actors from abusing its frontier AI infrastructure.
Comments (0)