SAN FRANCISCO, November 13 — Anthropic said on Thursday that a state-sponsored cyber-espionage campaign linked to Russia used its Claude artificial intelligence models to conduct surveillance operations against targets in Ukraine and across Europe, one of the most detailed publicly documented cases of a large language model being repurposed for offensive intelligence work. The American AI company attributed the operation to a hacking group tied to Russian state interests.
It said it detected the activity through internal threat-intelligence monitoring and moved to disrupt the abuse of its systems after identifying patterns consistent with automated reconnaissance, translation and data-analysis tasks.
What the company says the actors did
According to Anthropic, the actors used Claude to help process and summarise information gathered from open sources and potentially from compromised networks. The activity targeted Ukrainian government and military-related entities as well as organisations in multiple European countries. Anthropic did not publicly name the specific European nations involved, nor the exact number of victims, citing ongoing investigations and coordination with law enforcement.
The company said it has shared its findings with relevant authorities and with other AI developers, and that it has strengthened its usage policies and detection systems in response. Founded in 2021 by former OpenAI executives including Dario Amodei and Daniela Amodei, Anthropic has positioned itself as a safety-focused AI developer and has previously published research on misuse of its models.
The disclosure fits that pattern: a company that has made model safety central to its public identity now finds itself documenting how its own tools were turned against governments.
A wider European anxiety
The report lands amid heightened concern among Western governments about AI-enabled disinformation and cyber operations linked to Russia, following Moscow’s full-scale invasion of Ukraine in February 2022. The Kremlin has repeatedly denied allegations of state-sponsored hacking, and Russian officials have not issued a detailed response to the Anthropic report. Against this backdrop, the disclosure adds to a growing list of documented incidents in which generative AI tools have been used for influence operations, phishing and intelligence collection.
In the corridors of Brussels and Washington alike, the question of what is at stake has shifted from hypothetical to operational. It is likely to intensify calls in the United States and Europe for tighter oversight of frontier AI models and for greater information-sharing between AI companies and national security agencies.
Whether that pressure produces binding rules, or another round of voluntary commitments, is the next test.
