# Claude misused (2/7) : One consultant can now build a surveillance state > A single contractor in Bamako used Claude to engineer a system covering roughly 25 million SIM cards. Anthropic banned the account, and says the ban did nothing to the system itself. - Author: Abhay Bagda (https://abhaybagda.com/about/) - Published: 2026-10-15 - Topic: Quick Take - Series: Part two of seven - Reading time: 5 min - Canonical URL: https://abhaybagda.com/writing/one-consultant-surveillance-state/ --- Building a system that listens to a whole country used to take a department, a vendor contract and a few years. In the surveillance section of Anthropic's September threat report, one of them was built by a consultant with a Claude subscription. This is part two of seven on that report, and [part one was hacking](/writing/good-hack-stopped-telling-who). Surveillance gets ten cases, nearly all of them state security bodies or the contractors who work for them. Anthropic's own summary of the trend is that AI "is now being used in place of an engineering workforce." I'm in Singapore, and two of the ten have Singapore in them. One banned account was building a platform to profile social media users in Iran and the Gulf, and Anthropic says it was run by or for a company called S2T Unlocking Cyberspace, which "open-source research suggests is an Israeli-Singaporean commercial intelligence vendor." It was caught at the pilot stage, with no evidence it had been used on real targets. Another, a China-based desk compiling dossiers on religious figures across Asia, had Christian missionary networks linked to Singapore on its list. Neither made my three, but after those two I couldn't read this section as somebody else's region. > Banning the account ends the engineering. It doesn't un-build what was engineered. **The first is Mali.** A consultant in Bamako, likely working with the state intelligence service, used Claude as what the report calls "the primary engineering workforce" for a platform named Lakana 360. It covers all three national mobile operators, roughly 25 million SIM cards in scope, and collects call records, text messages and voice. It can recognise a person by their voice across different SIM cards, which defeats the cheap trick of switching to a burner, and it flags people for using encryption or a VPN. Malian law requires a court order before certain records are handed over. At the operator's request, the warrant check was removed from the part that writes an intelligence dossier on any phone number. The finished system runs on local AI models inside Mali, so the ban didn't reach it. In Anthropic's words, enforcement "disrupted the actor's software and design activities, but not the deployment of the platform." **The second is a recruiter who couldn't speak the language.** A Chinese-language operator, assessed with low confidence to be a contractor for PRC state security, was tracking Uyghurs in Syria. They fed Claude chatter pulled from more than 100 WhatsApp groups and dozens of Telegram channels and had it build profiles of people who might be approachable because of money trouble, family separation or disillusionment, noting who still had relatives in Xinjiang. Then they ran a multi-day covert recruitment attempt in Syrian Arabic, a language the operator didn't have. Claude drafted the outreach, translated the replies as they came in, and played an Arabic-speaking "expert" to check whether the deception was convincing. It declined some of the worst requests, including covert interrogation and mass-producing fake personas, but it did the rest. **The third is a security bureau that wrote the manual.** A cluster of accounts linked to municipal police and state security bodies in China used Claude for what the party-state calls "stability maintenance". One local bureau ran a daily briefing pipeline on overseas dissidents, then wrote the workflow up as an internal AI manual for colleagues, including the prompt that tells Claude to play an intelligence analyst serving the state. The same bureau asked for the gathering point, route and end point of a pro-democracy march in Vancouver. In a linked case Claude refused to produce a weekly report, the user rephrased, and it then produced guidance naming ten private citizens for "control". Anthropic's verdict on itself is blunt: "Our existing safeguards did not perform uniformly in these cases." Jacob Klein, who runs threat intelligence at Anthropic, [put the whole section in two sentences](https://www.implicator.ai/anthropic-claude-surveillance-mali-china-iran/). "Authoritarian states are using AI for surveillance, repression and influence operations today," he said. "It's no longer hypothetical." What gets me is how little of this is clever. Mostly they asked for a database screen, a translation, a daily report template, and each of those looks like ordinary office software work. Anthropic says as much about an Iranian case in the same section, where Claude "refused explicit profiling and propaganda requests, but our safeguards did not refuse many of the surveillance software tooling requests." So the model says no to profiling a dissident and yes to building the case-management screen the dissident's file will live in, and it's the screen that lasts. If you build software for a living, don't skip this bit. The dangerous deliverable in these cases was a normal-looking system with one check deleted. When someone asks you to make an approval step optional, switch a control off by default, or keep records forever, ask who that's for and get the answer in writing. You won't always like the answer, and you can't un-ship it afterwards. The report also tells the rest of us, quietly, what these systems look for: a burner SIM doesn't hide your voice, and switching on a VPN is itself a flag. The surveillance section is [here](https://www.anthropic.com/threat-intelligence-report-september-2026#surveillance-operations-sep-26), and it isn't light reading, least of all the Mali pages. Part three is influence operations, where the case closest to home is Malaysia.