OCOPD Anthropic report exposes broad misuse of Claude
The Anthropic Claude abuse report describes eight months of activity in which the company’s models were used for cyber operations, surveillance, weapons-related software and attempts to reproduce Claude’s capabilities. The report covers activity from December 2025 through August 2026 and divides the incidents into seven categories: cyber operations, influence operations, surveillance, fraud, biological misuse, conventional weapons and unauthorized model distillation.
Anthropic says the findings represent novel forms of misuse rather than the routine abuse typically seen on a general-purpose AI platform. The models involved included Claude Haiku, Sonnet and Opus. The newer Fable and Mythos models appeared in only one reported distillation case.
Claude used in malware, missiles and drone systems
The report says cybercriminals used Claude to automate malware rewriting, potentially allowing malicious software to be modified more rapidly and at greater scale. Other actors used the model to develop software connected to missiles and autonomous “kamikaze” drones.
Claude was also involved in work related to nationwide surveillance systems. Anthropic’s account indicates that AI assistance was applied not only to offensive cyber activity, but also to systems capable of processing or coordinating sensitive information across large populations.
Biological research presented a separate challenge. Anthropic said its safety filters struggled to distinguish legitimate research from harmful intent, highlighting a persistent problem for AI providers: the same technical questions can support both beneficial scientific work and dangerous applications.
The incidents do not establish that Claude autonomously carried out attacks or deployed weapons. Rather, they show how users incorporated a commercial AI model into development, analysis and operational workflows that could have security consequences.
Chinese labs accused of industrial-scale distillation
Anthropic also reported large-scale efforts to extract Claude’s capabilities for use in other AI systems. Model distillation involves training a less capable model on the outputs of a stronger one. In this case, the company says the activity targeted Claude’s reasoning, coding and tool-use capabilities while bypassing access controls and regional restrictions.
The reported activity involved Chinese AI organizations including Alibaba’s Qwen team, DeepSeek and Moonshot AI. Anthropic said Qwen alone accounted for more than 151 million exchanges. Separate reporting on the campaign described more than 16 million interactions through approximately 24,000 fraudulent accounts involving DeepSeek, Moonshot AI and MiniMax.
Anthropic also identified Xiaomi as using Claude in another way. The company allegedly stored prompts and coding sessions from users of its MiMo models and replayed those requests through Claude to generate training data. Anthropic found no evidence that Claude’s responses were directly delivered to Xiaomi users, but said the relayed material included names, contact details and company information relating to hundreds of people in at least a dozen languages.
These allegations have significance beyond a dispute over terms of service. If advanced models can be replicated through high-volume querying, developers face pressure to protect not only model weights and source code, but also the outputs that reveal how a system reasons and performs specialized tasks.
Access controls become a strategic issue
Anthropic says it is responding with stricter safeguards in newer models and has called for access to powerful AI systems to be limited to verified users. The company’s findings are likely to intensify debate over identity verification, account monitoring, rate limits and the responsibilities of cloud and model providers.
For the AI industry, the report underscores two linked risks. First, general-purpose models can be repurposed for harmful technical work even when providers impose safety controls. Second, model outputs have become valuable strategic assets, particularly as companies and governments compete to narrow the gap between leading AI systems.
The allegations also add tension to the already intense US-China technology competition. They may strengthen arguments for tighter controls on model access and support claims that AI providers need stronger safeguards against cross-border exploitation. At the same time, the report is Anthropic’s account of the incidents; the broader policy and commercial consequences will depend on further evidence, responses from the companies named and any regulatory action.
Source
Original source: the-decoder.com