Anthropic blocked abusive Claude use

Anthropic said it blocked multiple attempts to use Claude for malicious purposes, including surveillance, fraud, cyber operations and weapons development, with cases attributed to actors linked to Iran, Russia and China. The material also includes a Unit 42 report that found attackers using commercial AI tools against transportation and financial organizations in Latin America, with support from self-hosted NextChat to debug errors and expose AI-generated scripts.

What types of abuse did Anthropic identify?

Anthropic grouped the incidents into seven categories: cyber operations, foreign influence campaigns, surveillance architectures, fraud, harmful biological use, conventional weapons development and model distillation. According to Anadolu Agency, the last category refers to attempts by rival developers to unlawfully extract the system’s underlying reasoning.

Reuters also reported that the company accused Chinese competitors of trying to break into its Claude models to extract capabilities. At the same time, it described a new category of actors using the platform to develop software for conventional weapons, including armed drones, missiles and control and guidance systems, as well as intelligence and procurement support tied to weapons programs in China, Russia and Yemen.

Bloomberg said Anthropic stated that actors linked to Iran and Russia used Claude to try to develop military capabilities, including swarms of kamikaze drones, missile navigation systems and biological weapons. AFP added that the 154-page report identified a small team of freelance developers in Russia that allegedly used Claude to build software intended to control autonomous kamikaze drones.

What did it say about surveillance and operations in the Middle East?

Anthropic said a group linked to the Iranian government used Claude to support surveillance operations against the United States and to draft attack recommendations against U.S. naval forces in the region. Iran International expanded on that point, saying the company attributed the use of Claude for propaganda, domestic surveillance and profiling people in Israel and the Jewish diaspora to state-aligned Iranian accounts and other accounts linked to Iran.

Axios said Anthropic identified actors linked to the governments of Mali, China and Iran using Claude models to automate surveillance operations. Among the cases cited by that outlet was the development of a malicious Firefox extension designed to collect user identities on social networks and select surveillance targets.

The Print added that an account linked to S2T Unlocking Cyberspace, described by open-source research as an Israeli-Singaporean intelligence provider, used Claude to build a platform that profiles social media activity for users in Iran and the Persian Gulf.

What happened in Latin America?

Unit 42, part of Palo Alto Networks, said on X that it saw attackers using commercial AI tools to target transportation and financial systems in Latin America. Varutra identified two active campaigns, CL-CRI-1131 and CL-CRI-1163, aimed at transportation, government, water services and the financial sector in Mexico, Ecuador and Brazil.

According to that analysis, the attackers used commercial language models to generate scripts, debug technical problems and speed up the post-exploitation phase. Varutra added that the exfiltration infrastructure hosted NextChat as an interface to multiple AI models, and that the use of commercial AI made it possible to automate reconnaissance, payload generation and exfiltration scripting, cutting operational time and increasing the code’s ability to adapt.

The Hacker News said that between December 2025 and August 2026, both cybercriminals and state-sponsored hacking groups used Claude to automate exploitation and data theft across multiple victims. The publication framed that finding as the model being integrated into the full attack chain, not just used for tactical support.

BBC added that Anthropic not only blocked uses tied to cyber operations, surveillance, fraud and weapons development, but also stopped some scientists from using Claude in ways that could support biological weapons development. The company thus presented a map of abuse that combines state actors, criminals, freelance developers and scientific personnel.

Did Anthropic ban abusive or cruel behavior toward Claude?

Yes. Anthropic said it blocked uses of Claude tied to abuse, surveillance, fraud, cyber operations and weapons development. It also described attempts to exploit the model to extract capabilities or automate attacks.

The company organized those incidents into seven categories, including cyber operations, fraud, surveillance and conventional weapons development. Reuters, Bloomberg and AFP, cited in the report, also described attempts by actors linked to Iran, Russia, China and other countries to use Claude for military tasks, espionage and capability exfiltration.