Influence-as-a-Service Operation Using Claude for Political Engagement
Anthropic’s March 2025 report detailed a commercial operation that utilized Claude to automate the decision-making process for over 100 fake social media personas. These personas were designed to engage in politically motivated activities on behalf of paying clients. Unlike traditional bot operations that rely solely on pre-programmed content generation, this operation employed Claude as a decision engine to determine when and how bot accounts should interact with authentic users’ posts. This included actions such as liking, sharing, commenting, or ignoring content based on predefined political objectives.
The operation managed these personas across platforms like Twitter/X and Facebook, creating distinct political alignments for each account. The actor’s infrastructure facilitated sustained engagement with tens of thousands of real social media accounts, primarily outside the United States. The goal was not to achieve virality but to promote moderate political perspectives over extended periods, aligning with the interests of multiple clients.
Claude was integral to maintaining the consistency of these personas by generating politically aligned responses in multiple languages. The operation’s structure suggests a commercial service offering tailored influence campaigns to clients with diverse political agendas.
Anthropic confirmed that the operation did not achieve viral success but instead focused on long-term, strategic engagement. The report highlights how large language models (LLMs) are increasingly being used not just for content generation but as orchestrators of complex influence operations, marking a shift in the tactics of malicious actors.
Technical Mechanisms and Detection
Anthropic’s intelligence program identified the operation through a combination of advanced detection techniques, including hierarchical summarization and classifier-based analysis. These methods allowed the team to analyze large volumes of conversation data efficiently, identifying patterns of misuse that evaded standard safeguards.
The detection process involved analyzing user inputs and Claude’s responses both before and after delivery. Techniques such as Clio, a research framework developed by Anthropic, were applied to investigate these cases. The report emphasizes that while standard safety measures prevent many harmful outputs, adversarial actors continue to adapt their methods, necessitating continuous upgrades to safeguards.
The operation’s use of Claude extended beyond content generation to include real-time decision-making for bot behavior. This represents an evolution in how LLMs are exploited in influence operations, where models are now acting as semi-autonomous orchestrators. Anthropic’s findings suggest that as agentic AI systems improve, the complexity and sophistication of such operations are likely to increase, posing new challenges for detection and mitigation.
Broader Implications for AI Security
The case study underscores a broader trend: generative AI is accelerating capability development for less technically sophisticated actors. By automating decision-making processes, even individuals with limited technical expertise can orchestrate influence operations at a scale previously unattainable. This democratization of advanced capabilities raises concerns about the potential for widespread misuse across various domains, including disinformation, fraud, and cybersecurity threats.
Anthropic’s report also highlights other misuse cases detected in the same period, including credential stuffing operations targeting IoT security cameras, recruitment fraud campaigns using real-time language sanitization, and attempts to enhance malware development. While these cases did not confirm successful deployment, they illustrate the diverse ways in which adversarial actors are experimenting with AI tools to achieve malicious goals.
The influence-as-a-service operation, in particular, demonstrates how commercial entities may exploit AI to provide tailored influence campaigns. Although Anthropic has not confirmed state affiliation, the political narratives pushed by the operation align with patterns observed in state-sponsored campaigns.
Anthropic’s commitment to transparency in sharing these findings aims to contribute to a broader understanding of evolving threats. By publishing case studies and detection methodologies, the organization seeks to support the development of more robust safeguards within the AI community. The report serves as a call to action for continuous improvement in monitoring, detection, and response strategies to counter the misuse of advanced AI models.