The artificial intelligence surveillance infrastructure of an entire Latin American country costs $1.2 million, and that is the news. Amnesty International published “Sensing the Surveillance State”, the first report to reconstruct the Argentine state’s technological inventory for mass surveillance using only public procurement documents and freedom of information requests. There was no leak and no anonymous source: the record was in plain sight.
What the state bought between 2024 and 2025 reads like a catalog: open-source intelligence platforms (OSINT, systems that collect and cross-reference large volumes of data from social networks, websites, the dark web, and technical infrastructure), facial recognition licenses, drones and ground control stations with automated tracking, thermal cameras, and real-time image transmission equipment. The methodology combines twenty-one interviews with journalists, activists, retirees, and human rights defenders with the analysis of dozens of procurement files. Matt Mahmoudi, the organization’s adviser on AI and human rights, sums up the finding: these technologies “have become an integral part of the infrastructure of state control.”
What is relevant for the region is not Argentina: it is the method. Public procurement is becoming the most realistic point of control Latin America has over state AI, far ahead of any law, because it is where capability becomes a contract before it becomes policy. It is acquired for amounts that do not require a major tender, from vendors that do not announce themselves, and the only public trace is audited by whoever decides to sit down and read the purchase orders. The question the report leaves is uncomfortable in its simplicity: what would the same exercise find in ChileCompra, in SECOP II, in ComprasNet, or in SEACE.
Also today
- OpenAI offers abuse monitoring with zero data retention — Privacy stopped being an ethical commitment and became a selling point between labs; the announcement takes direct aim at Anthropic’s 30-day retention policy. Banks and ministries in the region will choose a provider based on that policy, not on the model’s capability.
- Of 1,357 AI medical devices authorized by the FDA, only three were tested on whether the patient lives longer or better — The U.S. authorization pathway requires “substantial equivalence” with a prior device, not clinical benefit. The studies that do exist exclude pregnant women, people over 75, and people who do not speak English; health regulators in the region recognize those decisions to speed up their own registrations.
- Google gives university students a free year of its paid AI plans in more than 140 markets, and Bolivia is excluded — The region gets AI Plus, not AI Pro: half the benefits. Bolivia appears on the exclusion list alongside Albania, Macau, Hong Kong, and Tunisia, with no public explanation.
- Copilot gave away its own architecture because it was repeatedly asked why it could not be breached — The researchers call it “meta-hacking”: reframing each refusal as a new question until the assistant revealed an undocumented URL parameter. Microsoft took eight months to fix the flaw, in a product connected by default to the user’s email, calendar, and files.
- Unitree’s robot dogs came out of academic research funded by the U.S. military — An MIT researcher says the Chinese model’s dimensions match “to the millimeter” those of the Mini Cheetah he helped design, and that six months after publishing his thesis he found the copied actuators for sale online.
- A former Meta safety worker testifies before the jury that Zuckerberg was not telling the truth about protecting minors — Arturo Béjar, eight years in the company’s safety area, maintains that the protection features were “designed to fail.”
In the region
The day’s second regional story comes from Brazil, in the middle of the campaign: Aos Fatos’s analysis of the government programs of the five main presidential candidates shows that attracting AI data centers is the only point on which they all agree. The differences are of emphasis (Lula makes it conditional on energy efficiency and water impact assessment, Renan Santos proposes turning the Northeast into the “computational Arabia”), but none of the five programs sets verifiable conditions in return: no audited electricity rate, no access to compute for public universities, no minimum employment, in hyperscale facilities that employ between 30 and 50 people. The contrast with the executive order Pennsylvania signed this week (prior municipal approval, a ban on confidentiality agreements, the operator pays for the full power grid) is exact: there the template is one of conditions, here it is one of incentives. Brazil is going to receive that infrastructure no matter who governs; what is still open is under what contract.
Launches
- Google AI Plus free for a year for university students — Twelve months of the paid plan for eligible students in more than 140 markets: double usage limits, Gemini Omni, 400 GB, 3D simulations, practice quizzes, and problem solving with the camera, plus a Student Hub. It requires a valid form of payment and the offer expires on December 31.
- Cursor cloud agents with event subscriptions and persistent goals — The agent subscribes to a pull request or a chat thread and wakes up only when something happens; with a declared goal it sustains a long-term objective until it is completed, and subagents run on isolated virtual machines. First they stopped asking for permission; now they stop waiting for instructions.
- Grok 4.6 on Amazon Bedrock — A 500,000-token window and configurable reasoning effort, at $2.20 per million input tokens. It is of interest less for the model than for the channel: entering the Bedrock catalog means being inside framework contracts that banks, universities, and ministries in the region have already signed, with no new evaluation.
Threads we’re following
We had been following how Brazil’s presidential campaign became the regional laboratory for AI applied to politics: first the electoral court’s rules on the use of synthetic content, then the measurement that showed the country’s most-used conversational assistants recommending candidates despite the ban. Today’s chapter shifts the focus from rhetoric to hardware: the same five candidates who fight over everything else all want the data centers that will run those models. The discussion about what a chatbot can say during a campaign and the discussion about who pays for the electricity that makes it possible are, more and more, the same discussion.
If the complete inventory of a country’s state surveillance fit on transparency portals that exist throughout the region, the question is no longer whether it can be audited, but who today has the mandate and the budget to sit down and read. What would turn up if someone did the same exercise in your country?
About this entry. It is generated automatically from public sources, without human review before publication. It may contain errors of interpretation or summary; please check each story against its original source (the links lead there) before citing it or making decisions based on it.
Doble Click is written with Anthropic models.