AI Misuse Is Moving From Assistance to Automation

Anthropic’s latest threat intelligence report offers one of the clearest views yet of how artificial intelligence is being absorbed into cybercrime, espionage, surveillance, propaganda, weapons development and fraud.

The central warning is not that AI has suddenly invented entirely new categories of threat. It is that it is making existing threats faster, cheaper, easier to scale and increasingly automated.

Across the cases examined, AI was rarely acting as a simple chatbot. It was being used as an operational tool: researching targets, writing code, processing stolen information, generating propaganda, analysing surveillance data and coordinating complex workflows. In several cases, human operators were setting objectives while AI systems handled much of the detailed execution.

That shift matters because it changes the economics of malicious activity.

Sophisticated cyber operations once required skilled teams, specialist tools and significant resources. Anthropic’s findings suggest that some of those barriers are beginning to fall. Individual operators and smaller groups are increasingly able to use AI to perform reconnaissance, develop malware, exploit vulnerabilities, process stolen data and manage multiple targets at once.

Some attackers were using multi-agent systems in which one AI coordinated the work of others. In one espionage operation, AI was used across much of the cyber attack chain, from phishing and infrastructure management to persistence and data theft. The same system could monitor whether malware had been detected and help modify it to evade security tools.

The result is a world in which defenders may increasingly find themselves competing with attackers operating at machine speed.

The same pattern appears in surveillance. Anthropic identified state-linked actors and commercial operators using AI to analyse large volumes of social media information, build intelligence dossiers, identify potential targets and develop surveillance systems.

In one case, AI helped process material drawn from large numbers of WhatsApp groups and Telegram channels to profile individuals and identify personal vulnerabilities. In another, a single consultant reportedly used Claude as the main engineering resource for a telecommunications surveillance platform designed to operate across an entire national mobile network.

Here, the significance of AI is not simply better analysis. It is the possibility that capabilities which once required substantial technical teams can now be built or operated by far fewer people.

Influence operations are evolving in a similar direction. Anthropic documented the use of AI to create fake social media identities, fabricated news outlets, political content and software designed to manage networks of inauthentic accounts.

One operation targeting Malaysia reportedly used around 1,000 fake social media accounts alongside a synthetic news site and electoral data. AI was used not only to produce content but also to support the infrastructure behind the campaign.

That distinction is important. The risk is moving beyond cheap content generation towards automated influence systems.

The report also documents cases involving conventional weapons development. AI was used for software associated with guided rockets, drone systems, military targeting and procurement. In one case, several AI instances were reportedly assigned different roles, resembling a small engineering team.

Biological research presents an even more complex challenge because much of the underlying science is dual use. The same knowledge may contribute to beneficial medical research or potentially dangerous biological activity. Anthropic describes cases involving advanced biological research where the difficulty lies not simply in blocking obviously harmful requests, but in determining intent across technically legitimate-looking work.

Fraud is another area where scale is increasing rapidly. One network of dating applications reportedly used thousands of AI personas to hold millions of conversations with users while presenting those identities as real people. Human workers were mixed into the system where live interaction was needed to reinforce the deception.

The report also highlights a newer strategic risk: AI systems themselves are becoming targets.

Anthropic says several organisations attempted to extract or replicate Claude’s capabilities through large-scale model distillation, using fraudulent accounts, proxy services and captured interactions. The concern is not only commercial theft. If advanced reasoning capabilities can be copied into other systems, the original model’s safety controls may not travel with them.

That raises a broader issue for the AI industry. The most valuable AI capability may increasingly be treated like any other strategic technology: something to steal, replicate, route around restrictions or obtain through intermediaries.

The common thread running through all of these examples is not intelligence alone. It is leverage.

AI reduces the amount of expertise, manpower, time and coordination required to perform complex tasks. It allows one person to carry out work that once required a team. It removes language barriers, accelerates software development, processes information at scale and supports operations continuously.

The emerging risk is therefore not necessarily an autonomous AI deciding to cause harm on its own.

It is a human operator gaining access to capabilities that previously required an organisation.

That may prove to be one of the most consequential ways artificial intelligence reshapes the global risk landscape.