In a new report released by global artificial intelligence company Anthropic said that it has uncovered a new wave of attempts wherein its Claude AI models were misused from cyberattacks and online fraud to surveillance and weapons-related activities, highlighting how quickly AI misuse is expanding beyond simple chatbot abuse.
The AI company said its Threat Intelligence team disrupted operations involving suspected state-backed groups, cybercriminals, commercial spyware vendors and politically motivated actors between December 2025 and August 2026.
In its latest report, Anthropic details cases spanning seven areas of harm: cyber operations, influence operations, surveillance, scams and fraud, biological misuse, conventional weapons development and AI model distillation.
The company said these are not typical examples of policy violations, but some of the most significant and unusual cases its security teams have detected. Anthropic said it blocked the activity, used the findings to strengthen its safeguards and shared intelligence with authorities and industry partners where appropriate.
One of the cases involved a network of fake dating applications created to defraud users.
The operation shows how AI can be incorporated into online fraud at scale. Dating scams typically depend on creating convincing identities and maintaining lengthy conversations with victims. AI can make parts of that process faster and easier to automate.
Anthropic included the case as an example of how organized groups are experimenting with AI to support fraudulent operations.
Another case involved surveillance systems designed to identify and monitor dissidents.
The example highlights concerns about AI being combined with large datasets and other software systems to analyse information about individuals. Anthropic has previously placed restrictions on certain surveillance, tracking and profiling uses of its models.
Cyber operations continue to be one of the biggest areas of AI misuse.
Anthropic has previously disclosed cases involving ransomware development, data theft and other cyberattacks. In one case, a threat actor used Claude to help develop ransomware and attempted to make the malware available to other criminals.
The company has also identified operations where AI was used across several stages of cyberattacks, including reconnaissance, analysing stolen information and preparing extortion demands.
The concern goes beyond AI simply generating malicious code. More capable models can potentially assist with multiple connected tasks, allowing attackers to automate parts of an operation that would previously have required several people with different technical skills.
Anthropic has also identified attempts to use Claude to conduct AI model distillation—the process of extracting the capabilities of one AI system through large numbers of queries and using the responses to develop another model.
In February 2026, Anthropic said it detected large-scale distillation campaigns involving DeepSeek, Moonshot AI and MiniMax. Around 24,000 fraudulent accounts were involved, generating more than 16 million exchanges with Claude, revealed the company.
These activities went beyond normal experimentation and represented systematic attempts to extract capabilities from its models, said the company.
The latest report also covers cases involving influence operations, biological misuse and conventional weapons development.
The actors identified by Anthropic include suspected state-sponsored groups, financially motivated criminals, commercial spyware vendors, state propaganda organisations and politically motivated individuals.
Together, the cases point to a broader shift in how advanced AI is being tested and used. The technology is no longer being targeted only for generating harmful content; threat actors are increasingly trying to incorporate AI into wider, coordinated operations.
Anthropic said threat actors continue to test its safeguards and look for ways around its detection systems.
For AI companies, that makes safety a moving target. As models become better at coding, reasoning, analysing information and carrying out complex tasks, developers also need stronger systems to detect when those capabilities are being turned toward malicious activity.
Anthropic said it is sharing details of the cases to help other AI developers identify similar patterns and help governments and security organisations prepare for emerging threats.
The latest report offers a clear warning for the wider AI industry: the race to build more capable models is also creating a race to prevent those capabilities from being exploited.
