Anthropic Reveals cinematical AI Misuse in Government Surveillance

Anthropic Reveals AI Misuse in Government Surveillance: How Claude Was Used to Build Monitoring Systems

Artificial intelligence is increasingly being used to write code, analyse information, automate business operations and assist with research. But a new report from Anthropic highlights a more troubling application: the use of AI tools to support surveillance operations.

In its September 2026 threat intelligence report, Anthropic disclosed cases in which state-aligned actors, government-linked contractors and commercial surveillance operators misused Claude. The reported activities involved building surveillance software, collecting information, analysing social media activity and identifying people who could become targets for monitoring.

The cases involved actors linked to Mali, China and Iran, alongside commercial surveillance activity involving people in Iran and the Persian Gulf region.

Anthropic says it identified and disrupted the activities, banned associated accounts and improved its safeguards.

The report raises an important question: when an AI assistant can help build software and analyse enormous amounts of information, how easily can that capability be turned into a tool for monitoring people?

1. What did Anthropic discover?

Anthropic's Threat Intelligence team investigated misuse of Claude between January and July 2026.

The September report describes several surveillance-related operations involving state-aligned actors, state-linked contractors and commercial spyware vendors.

The reported activities were not limited to asking an AI chatbot questions about people. Some operators used Claude as a software development assistant to create surveillance infrastructure. Others used it to process social media information, classify individuals and prepare intelligence-style reports.

Anthropic identified several recurring patterns:

- AI was used to help build software that supported surveillance.

- Large collections of social media information were processed in bulk.

- Individuals were classified using characteristics such as location, demographics and political views.

- Some operations used AI to prepare reports and identify potential targets.

- AI-assisted tools were being integrated into existing security and intelligence workflows.

This is important because it shows how AI can become part of an operational system rather than remaining a simple question-and-answer assistant.

Anthropic says the activity violated its usage policies, which prohibit non-consensual surveillance and profiling that violates civil liberties and human rights.

2. Mali: Claude helped build a mass-surveillance platform

One of the most striking cases involved a consultant working with Mali's national security authorities.

According to Anthropic, a single subscriber used Claude as the primary engineering workforce to develop a surveillance platform called Lakana 360.

The system was designed to collect and process telecommunications information across Mali's three national mobile operators. Anthropic estimated that the platform was designed to cover approximately 25 million SIM cards.

The reported capabilities included collecting call records, text messages and voice communications, as well as generating intelligence dossiers associated with phone numbers.

The platform was also designed with additional surveillance capabilities, including identifying users across SIM cards, monitoring geographic areas and matching information against other government registries.

A particularly concerning detail was the handling of legal safeguards. Anthropic reported that a component originally designed to require a court order before certain information could be disclosed was modified at the operator's request so that the requirement was removed.

This case illustrates a distinction that is important to understand: Claude was not simply being used to read surveillance dossiers. Anthropic says it was used to engineer the underlying software that enabled the intelligence-gathering system.

The platform demonstrates how an AI coding assistant can potentially reduce the amount of specialist engineering work needed to develop complex monitoring infrastructure.

Anthropic disrupted the account involved. However, the report also describes how the platform was subsequently deployed locally using other models after access to Claude was cut off.

That last detail matters. Blocking misuse on one AI platform does not necessarily eliminate the wider surveillance operation.

3. China: AI-assisted analysis of social media and potential targets

Anthropic also described multiple operations involving China-linked actors.

In one case, a religious affairs intelligence collection unit reportedly used an AI assistant to produce thousands of investigations per month.

Anthropic said an office that had previously involved many teams of analysts had been reduced to a single office using AI to support its work.

Other cases involved analysing social media posts and news articles to identify political sensitivity and potential targets for what the actors described as "control."

The report describes operations in which AI was used to help prepare intelligence briefings concerning overseas dissidents and civil society organisations.

It also describes an operation involving the monitoring of individuals and communities outside China.

These findings are significant because they show how AI-assisted analysis can be integrated into existing bureaucratic processes.

A large volume of public information can be difficult for a human team to examine manually. AI systems can help organise that information, classify it and produce structured summaries.

But the ability to process information does not make the resulting classifications accurate or fair.

An AI system might incorrectly interpret a person's political views, misunderstand context or associate someone with an activity they did not participate in.

When such assessments are used in real-world monitoring, errors can have serious consequences.

4. Iran: Surveillance tools and social media profiling

The report also describes several Iran-linked operations.

In one case, Anthropic identified accounts associated with two linked units that allegedly used Claude to develop surveillance tools and support a central case-management system.

Anthropic reported that these actors used Claude to build a malicious Firefox browser extension designed to harvest user identities from social media platforms.

The report also describes the use of social network analysis to examine large volumes of posts and identify opposition-related accounts.

One reported analysis involved 155,216 tweets. Another workflow identified 39 Iranian opposition and diaspora accounts for monitoring.

Anthropic said it found evidence linking the units to Iranian paramilitary domestic security entities.

The report also describes how information was collected and organised in a central system containing personal details, social accounts and other information about individuals.

Anthropic banned 16 accounts associated with the two units.

The significance is not merely that an AI model helped write software. It is that the software was reportedly connected to a broader surveillance workflow.

5. How AI can change surveillance operations

To understand why these findings matter, consider the difference between traditional manual monitoring and AI-assisted information processing.

In a conventional workflow, a team might collect information, review it, classify it, prepare reports and identify people for further investigation.

Each stage requires time and human resources.

AI can help automate parts of this process.

At a high level, an AI-assisted surveillance workflow might involve:

Information collection: Information is gathered from sources accessible to the operator, such as social media or other datasets.

Data organisation: AI helps organise large amounts of unstructured information into structured records.

Pattern analysis: Systems identify patterns, connections or changes in the collected information.

Classification: People or records may be grouped according to selected characteristics.

Human review or operational decisions: Investigators may review the results and decide what action to take.

This is a high-level explanation, not a recommendation for building surveillance systems.

The central concern is that automation can make large-scale monitoring cheaper and easier to operate.

A small team may be able to perform tasks that previously required a much larger group of analysts.

However, AI-generated classifications are not automatically reliable. The quality of the output depends on the data, the system and the way its results are interpreted.

6. Why AI-powered surveillance raises privacy concerns

Surveillance is not automatically illegal or harmful in every situation. Governments may use lawful monitoring for legitimate public-safety purposes.

But surveillance becomes particularly concerning when it is conducted without meaningful legal safeguards, transparency, accountability or respect for individual rights.

AI introduces additional challenges.

Privacy: Personal information can be collected, combined and analysed at a scale that individuals may not expect.

Misidentification: AI systems can make incorrect connections between people, accounts and activities.

Political profiling: Analysing someone's posts or online associations can lead to assumptions about their political beliefs.

Lack of transparency: People may not know that an automated system has classified them or contributed to a decision affecting them.

Limited ability to challenge decisions: If an AI-generated report is treated as authoritative, an individual may have difficulty correcting inaccurate information.

Scale: Automation can make it easier to monitor far more people than a traditional manual process could support.

These concerns are especially relevant to journalists, activists, minority communities and people living outside their country of origin.

The broader issue is whether people retain meaningful rights when automated systems are used to collect and interpret information about them.

7. What did Anthropic do after discovering the misuse?

Anthropic says it banned the accounts associated with the operations described in its report.

It also strengthened detection systems using the tactics and behaviour discovered during its investigations.

The company said it shared relevant indicators and intelligence with industry partners and authorities where appropriate.

These actions are part of a wider challenge facing AI developers.

AI companies need to prevent their products from being used for harmful activities while continuing to support legitimate software development, security research and information analysis.

That is not a simple technical problem.

A request to analyse public information may be legitimate in one context and part of an abusive surveillance operation in another. Detecting misuse requires more than identifying individual keywords. It can require understanding patterns of behaviour across multiple interactions.

There is also a limitation: banning an account does not necessarily stop the operator from switching to another AI provider, using open-source models or developing its own software.

Anthropic's report describes this wider challenge, including cases where operators continued their activities using other models after losing access to Claude.

This means platform-level safeguards are important, but they cannot solve the entire problem alone.

Governments, technology companies, regulators and civil society organisations all have roles in addressing misuse.

8. What does this mean for ordinary internet users?

The report should not lead people to believe that every AI company or government is using AI to monitor every individual.

That would be an unsupported conclusion.

But it does show why personal data protection matters.

People should be careful about what information they publish publicly, which applications they grant access to their accounts and what permissions they give browser extensions.

Businesses should also review how customer information is collected, stored and processed by AI services.

Developers building AI-powered applications should consider privacy, access controls, data retention and audit logs from the beginning.

For India, this discussion is relevant as businesses, public institutions and startups increasingly explore AI-based services.

The opportunity is to use AI for healthcare, education, productivity, customer support and public services while maintaining appropriate safeguards for personal information and civil liberties.

AI-powered analysis can be useful. But using technology to profile people without proper safeguards can create serious risks.

Conclusion: AI surveillance is no longer just a theoretical concern

Anthropic's September 2026 report provides concrete examples of how its AI assistant was misused in surveillance-related operations.

The cases show different forms of misuse: building telecommunications monitoring infrastructure in Mali, supporting social media analysis and intelligence workflows in China, and developing surveillance software in Iran.

They also show that AI does not need to be an entirely autonomous system to create significant consequences. Even when a human operator remains in control, AI can help automate parts of a workflow and make complex tasks easier to perform.

Anthropic says it disrupted the accounts it identified and strengthened its safeguards. But the wider challenge remains because operators can potentially move between platforms and models.

The important question is not simply whether AI can analyse information. It is whether that capability is being used with appropriate legal protections, transparency and accountability.

As AI becomes more powerful, where should society draw the line between legitimate security operations and invasive surveillance?