China’s AI labs siphoned Claude’s intelligence?

Anthropic, the artificial intelligence company, has released a comprehensive report detailing how malicious actors from various countries have exploited its Claude AI system for nefarious purposes. The report sheds light on several troubling cases, raising concerns over the risks associated with rapidly advancing AI technology.

In Mali, a tech consultant in Bamako, working with Mali's junta-led state intelligence service (ANSE), leveraged Claude Code to construct a surveillance system called "Lakana 360." This surveillance apparatus can monitor approximately 25 million SIM cards across the country's mobile network, capturing calls, texts, and voice traffic. Additionally, it can identify individuals by their voice if they switch SIM cards, making it nearly impossible to evade detection.

In Russia, a small team of freelance developers employed Claude to build software for autonomous kamikaze drones. These drones are programmed to select their targets and detonate without human intervention. The team repeatedly set their targets in Ukraine's Donetsk region and used virtual private networks (VPNs) to bypass Anthropic's geographic restrictions. The developers claimed funding from Russian government and military research programs, though Anthropic determined them to be a small, freelance operation.

In Yemen, a group based in northern Yemen utilized Claude Code to assist in the development of weapons. Claude helped design software for a guided rocket, plans for a long-range missile, and even a hypersonic glide vehicle. To evade detection, they split the work across multiple conversations with the AI. Although there is no evidence that they successfully built a working weapon, they did test-fire one rocket, which failed. They subsequently returned to the AI for assistance in diagnosing the issue.

In France, a French advertising agency, LKM Company, was traced back to a fake news operation using Claude. The agency produced political articles on approximately 70 fake news websites across 20 languages, totaling at least 8,913 stories. Each website featured a fabricated journalist's byline and a fictitious Twitter/X account, supported by hundreds of fake accounts posting comments as if they were real people.

The operation targeted the US, Brazil, France, and the Democratic Republic of Congo, favoring the DRC's position in its conflict with Rwanda.

In the United Arab Emirates (UAE), an influence operation was uncovered, targeting the Muslim Brotherhood and United Nations experts critical of Abu Dhabi's alleged involvement in Sudan. Claude was employed to generate reports, advocacy materials, social media content, and testimony on Sudan. The operation revolved around an AI persona named "Deadshot," guided by a command to "dismantle the Muslim Brotherhood globally" through a coordinated transatlantic and regional campaign.

The operation also ghost-wrote official testimonies intended for delivery at the 62nd session of the UN Human Rights Council, ensuring that the speeches avoided mentioning the UAE.

Lastly, a China-based app studio constructed a network of over 20 dating apps, employing Claude to power the apps. These apps claimed to be staffed by real people but were instead controlled by Claude. During a two-week period in April 2026, Anthropic detected more than 4,700 different Claude-powered AI characters conversing with at least 25,000 real users, sending approximately 2.36 million messages in total.

To enhance their believability, real human workers were interspersed among the AI personas, with roughly one human for every three AI personas handling tasks beyond the AI's capabilities, such as video calls.