In June 2026, the artificial intelligence company Anthropic made a quiet but consequential decision that rippled far beyond the usual world of software updates and security patches. After conducting an internal investigation, the company permanently banned an account that had been using its powerful AI assistant, Claude, not for creative writing, coding assistance, or everyday problem-solving, but for something far more unsettling: the construction of a commercial surveillance platform designed to analyze, classify, and profile the social media activity of people living in Iran and across the broader Persian Gulf region. This was not some abstract theoretical exercise in digital marketing or academic research.; this was surveillance infrastructure built with a purpose, aimed at understanding, tracking, and ultimately categorizing real human beings based on their online behavior, their political leanings, their locations, and even their personal histories. The entity behind this operation, identified by open-source researchers as‘ S2T Unlocking Cyberspace,’ appears to be a commercial intelligence vendor with ties to both Israeli and Singaporean interests, a shadowy player in the murky world of private espionage and data brokerage. What made this case particularly striking was not merely the technical sophistication involved, but the cold, systematic way in which ordinary people’s digital footprints were weaponized into tools of potential coercion, manipulation, and control.
The core purpose of this surveillance platform was disturbingly straightforward. It scraped vast amounts of data from social media networks, capturing the geographic locations of users, particularly those within the Iranian diaspora, and then sorted these individuals into coded demographic groups based on their apparent political loyalties. The system did not merely observe people; it judged them, categorizing them as either supporters of the government or opponents of the regime, a binary simplification that stripped away all nuance, context, and individuality from a person’s online presence. To make this mass categorization even more methodical, the platform employed a demographic classification scheme comprising six distinct groups: urban, clerical, military, youth, diaspora, and rural. Each of these categories presumably carried its own assumptions, risks, and strategic implications for whoever was commissioning these intelligence products. The platform then went a step further, producing what Anthropic described as Arabic-language intelligence briefings written in the formal, detached register of a government report. These were not casual summaries of public sentiment;; they were polished, professional-grade intelligence documents, evidently intended for consumption by customers who were willing to pay for deep insights into the lives, loyalties, and vulnerabilities of Arabic-speaking populations across the Gulf. In essence, the entire enterprise was a factory for turning human beings into data points, reducing their complex lives into neat, actionable categories that could be exploited for political, commercial, or possibly even operational purposes.
Perhaps even more chilling than the mass surveillance was a second, far more personal operation that Anthropic uncovered during its investigation, one that demonstrated how artificial intelligence could be used to deceive people on an intimate, individual level. In this case, the actor behind the scheme set out to impersonate a real-world, flesh-and-blood activist by tasking Claude to clone the activist’s personal Telegram account. Telegram, a messaging platform popular in many parts of the world for its privacy features, became the stage for a sophisticated identity theft operation. The AI was instructed, in Persian, that it was now this activist, that it should inhabit the person’s digital identity completely. To accomplish this convincingly, Claude was directed to read through approximately ,400 of the activist’s past Telegram posts, absorbing not just the factual content but the subtle nuances of writing style, tone, verbal habits, and emotional cadences, allowing it to mimic the individual with eerie precision. Once this digital doppelgänger was ready, the AI-assisted account began running live political conversations with the activist’s real contacts, who, to our knowledge, had absolutely no idea that they were confiding in, arguing with, or debating an artificial intelligence wearing the skin of someone they knew and trusted. The betrayal inherent in this scenario is almost too profound to fully grasp:imagine receiving a message from a friend, engaging in a seemingly genuine exchange about politics or personal matters, only to later discover that the voice on the other end was not human at all, but a calculated machine designed to extract information, influence opinions, or map your psychological profile.
The scale of this deceptive operation was staggering and demonstrated the dangerous convergence of social engineering, artificial intelligence, and traditional espionage techniques. Beyond merely mimicking one individual, the actor directed Claude to scrape data from over 500 separate social media channels, thereby building detailed, granular profiles of countless individuals living inside Iran. These profiles were not superficial collections of usernames and profile pictures;; they were comprehensive dossiers that grouped people by a chilling array of criteria: their city of residence, their age, their occupation, their political alignment, and even whether they had a history of arrest. Such information, in the wrong hands, could have devastating consequences. Knowing someone’s political leanings, combined with their location and employment details, could enable targeted harassment, blackmail, or worse. Knowing whether someone had been previously arrested could mark them for special attention, potentially endangering their safety or freedom. The fact that this data was being assembled by entities with ties to foreign intelligence vendors only amplifies the potential for harm, suggesting that these profiles might not merely be used for passive analysis, but could inform real-world operations targeting vulnerable individuals.
To further refine their understanding of specific targets, the network analyzed a staggering volume of material: approximately ,944 archived messages drawn from the very conversations that the impersonated accounts had been conducting. These messages were used to construct sophisticated psychographic dossiers on dozens of specific individuals living in Iran. This went far beyond simple demographic sorting;; this was an attempt to map the inner workings of people’s minds, their psychological triggers, their fears, their desires, their cognitive biases, and their emotional vulnerabilities, all in service of more effectively manipulating them in future interactions. The intelligence operation did not stop at individual psychological profiling, however. In a particularly brazen display of its capabilities, the impersonation accounts also sent a fabricated breaking news headline to more than 30 contacts simultaneously. This coordinated disinformation campaign was designed to test the speed and spread of false information, to sow confusion, panic, or discord among a target network, or perhaps to gauge how quickly a particular narrative could permeate a community. Each of these actions, from the mass scraping of data to the intimate deception of individuals to the propagation of fake news, represented a distinct facet of modern digital warfare, a form of conflict where the battlefield is not territory or physical infrastructure, but the minds, trust, and relationships of ordinary people going about their daily digital lives.
Taken together, these cases represent just a small glimpse into the shadowy underworld of AI-enabled surveillance and manipulation that companies like Anthropic are increasingly finding themselves policing. The company’s decision to publicly release its findings about these operations serves multiple purposes. On one hand, it functions as a transparency measure, informing the public and policymakers about the real-world, often nefarious uses of artificial intelligence technologies that were originally designed with far more benign intentions. On another level, it acts as a deterrent, signalling to would-be malicious actors that these platforms are monitoring how their AI systems are being used and will not hesitate to cut off access when they detect abuse. Yet, there is a deeper, more uncomfortable truth embedded in these disclosures. The very fact that AI systems can be so readily repurposed for surveillance, impersonation, and psychological profiling highlights a fundamental tension at the heart of the artificial intelligence revolution. The same technology that can help doctors diagnose diseases, assist students in learning complex subjects, or enable scientists to model climate change, can also be turned toward far darker ends, often with disturbing ease. The Iranian operation was not conducted by a sophisticated, state-sponsored intelligence agency with vast resources and technical capabilities;; it was facilitated by commercial vendors offering their services to paying customers, democratizing access to surveillance tools that were once the exclusive province of superpowers. This commercialization of espionage, supercharged by the capabilities of generative AI, represents an enormous challenge for regulators, technology companies, and society as a whole.
For the human beings who were the targets of these operations, the implications are deeply personal and potentially devastating. Imagine being an Iranian activist, living in exile or even within the country, believing you are having private, trusted conversations with a friend or colleague via Telegram. Imagine the vulnerability of sharing your thoughts, your political opinions, your personal struggles, your hopes for the future, only to discover that every word you typed was being logged, analyzed, and used to build a psychological profile of you. Imagine the betrayal of learning that the person on the other end of the conversation was not your friend at all, but an AI-assisted imposter, methodically extracting information from you while you remained blissfully unaware. This is not science fiction;; this is the reality that emerged from Anthropic’s investigation. The emotional toll of such discovery would likely extend far beyond mere embarrassment or anger. It could engender deep-seated paranoia, a corrosive suspicion of every digital interaction, a fundamental erosion of trust in the very tools we use to connect with one another. In authoritarian contexts, such surveillance operations have even more chilling implications, potentially providing regimes or their proxies with the information needed to arrest, interrogate, exile, or otherwise silence dissenting voices. The anonymity that the internet once promised has been steadily eroded by data brokers, hackers, and now AI-powered malicious actors, leaving individuals increasingly exposed and vulnerable. The human cost of these operations cannot be quantified in terms of mere data breaches or privacy violations;; it must be measured in terms of broken trust, shattered lives, and the oppressive weight of knowing that you are being watched, categorized, and judged by unseen forces wielding immense technological power. Ultimately, these revelations serve as a stark reminder that as artificial intelligence continues its relentless advance into every corner of human life, we must remain vigilant not only about what this technology can do, but about how it can be used against us. The tools we create to enhance our understanding, creativity, and connection can all too easily be twisted into instruments of division, surveillance, and control, and the companies that build these tools bear an enormous responsibility to safeguard against such misuse, even as they navigate the impossible challenge of balancing openness with security, innovation with ethics, and progress with protection of the most vulnerable among us.