Welcome to the 15th edition of HRF’s AI for Individual Rights newsletter.
This week, Saudi Arabia’s state-linked AI company HUMAIN announced a partnership with French AI lab Mistral. When authoritarian regimes collaborate with leading AI companies from democracies, they may gain greater influence over how AI systems are developed and deployed.
Meanwhile, a different partnership offers a more hopeful sign.
US tech giant NVIDIA made a $6 billion dollar deal with US AI startup Poolside to build capable open-weight models. Open-weight models are ones whose underlying parameters are publicly available. That means anyone with sufficient hardware can download, modify, and run the models. If the investment pays off, it could give users cutting-edge, user-controlled AI developed in a democracy.
The Latest in AI for Repression
French AI Lab Mistral Partners with Saudi Arabia’s HUMAIN
Mistral, the biggest AI lab in Europe, announced a major partnership with state-linked Saudi AI company HUMAIN. These companies plan to expand AI infrastructure, develop AI models tailored to Saudi Arabia and the wider region, and bring AI solutions to regulated industry. They say the collaboration is meant to develop “sovereign AI,” which means Saudi organizations could use AI while keeping their data, models, and computing infrastructure under their own control. But what will that sovereignty look like in practice? Could these capabilities also strengthen Saudi Arabia’s capacity for AI-powered surveillance and censorship? It’s deeply concerning to think that one of Europe’s most prominent AI labs could help an authoritarian regime build greater control over advanced AI infrastructure.
In context: Mistral is one of the most prominent democracy-based companies producing open-weight models. Because open-weight models allow users to download, modify, and run them independently, they can greatly empower individuals. This partnership is not a promising signal for a future in which powerful, open-weight AI is developed and controlled primarily by democratic actors.
Kazakhstan’s Smart-City Investment with a UAE-Based Company
Kazakhstan is reportedly expanding its $190 million partnership with Presight AI, a company based in the United Arab Emirates (UAE) that sells surveillance technology. Presight has openly marketed tools for facial recognition, video analytics, and AI-assisted policing. Now, it will help Kazakhstan’s authoritarian regime build smart-city systems and state-controlled AI infrastructure. The project includes creating virtual models of the country’s capital city, Astana, and its biggest city, Almaty, that will combine real-time data from transport systems, environmental sensors, and video analytics. In effect, Kazakhstan and a company based in another authoritarian regime are working together to build AI infrastructure capable of tracking people, analyzing activity, and watching dissidents. It is a reminder that one regime’s surveillance tech can easily become another’s.
In context: Kazakhstan already uses AI surveillance in ways that can reach political dissidents. One political blogger reported being stopped by police after facial recognition identified him as a “civil activist.” Further expansion of AI-powered surveillance will only make dissent more dangerous.
Slovakia Finds Spying Capabilities in Russian-Made Traffic Cameras
Slovakia’s cybersecurity watchdog, National Security Authority, found alarming vulnerabilities in 279 AI-powered traffic cameras deployed across the country. Slovakia purchased the cameras through a Cyprus-based company, but investigators later found they were simply rebranded versions of Russian-made cameras. Even more frightening, a hidden function allowed 12 Russian phone numbers to execute code on the cameras, potentially giving them remote control. A separate vulnerability let any attacker who knew a camera’s IP address view its live feed without a password. The spread of AI-powered surveillance cameras like these could help authoritarian regimes like Russia not only monitor their own citizens, but also surveil people from afar.
Turkey Pours Billions into a Major AI Push
Turkey has launched a new national AI action plan aimed at making the country more technologically advanced and independent. The plan calls for at least $10 billion in investment through 2030 to expand AI education efforts, data centers, AI infrastructure, AI-powered public services, and Turkish-made AI models. Some of these investments could benefit citizens and businesses by improving public services, education, health care, and productivity across industries. However, the same investments could potentially power AI-enhanced surveillance, faster analysis of citizens’ personal data and activity, or domestic AI models designed to enforce regime propaganda.
In context: Turkey has used AI-powered facial recognition to identify protesters, explored AI tools to flag people of interest, and blocked AI-generated content after it insulted Turkish dictator Tayyip Erdoğan. Building AI at home could give the Erdoğan regime a new toolkit for crushing dissent.
Bangladesh Expands AI-Powered Surveillance that Automatically Fines Traffic Violations
Bangladesh has introduced an AI-powered traffic enforcement system in which over 1,400 AI-powered cameras automatically fine traffic violators. The cameras detect speeding, wrong-way driving, and other traffic violations. When a violation is detected, the system sends the details to the regime’s vehicle database, and the registered owner receives a fine by text message. Enforcing traffic violations can improve road safety, but the system sets a concerning precedent for automated punishment. If AI can automatically fine everyday drivers, it’s not hard to imagine a future where it identifies and drains the funds of protesters or opposition figures.
In context: Hybrid and fully authoritarian regimes often freeze bank accounts, seize assets, or financially cut off people they view as threatening their power. AI-powered fines give regimes a new way to financially punish dissent.
Recommended Content
HRIC Digital Rights Report - July 2026
HRF’s AI for Individual Rights Fund grantee Human Rights in China (HRIC) published its fourth Digital Rights Report, a monthly briefing on the developments shaping digital rights in China. In this edition, HRIC discusses how China is expanding its global AI influence and what that growing technological soft power could mean for human rights. It also covers China’s tightening restrictions on VPNs and leaked efforts to censor information across Chinese social media. Give it a read here or subscribe to HRIC’s Substack to learn more about China’s evolving repression.
The Latest in AI for Freedom
NVIDIA Makes a $6 Billion Deal to Build Frontier Open-Weight AI
US tech giant NVIDIA made a $6 billion dollar deal with Poolside, a US AI startup, for the purpose of developing highly capable open-weight models. Through the deal, NVIDIA will gain access to Poolside’s technology and bring over 100 of its engineers to the effort. One of Poolside’s founders said the deal was meant to secure a future where AI “would not be a closed technology controlled by few but one built by many out in the open.” This investment could have real promise for an open-weight model that individuals, not companies or governments, can run and control themselves.
In context: Today, the strongest open-weight models are developed in China, where researchers have found that models reproduce the Chinese Communist Party’s censorship and political biases. NVIDIA’s efforts could help the world rely less on open models built under authoritarian regimes.
Ornith Launches New Models, One That Runs Locally on Phones
DeepReinforce.ai, a US-based AI research team, has launched Ornith-1.5, a family of open-weight models. The release includes three model sizes, ranging from a massive model that requires expensive hardware to a compact model that can run locally on a phone. Ornith says the smallest model can perform many complex tasks nearly as well as models many times its size. That means dissidents and human rights defenders can chat with more capable AI on their phones, completely privately, anywhere, without the internet. Try running it locally on an iPhone or Android with Atomic Chat.
In context: Ornith was built on top of the Chinese open-weight model Qwen. This shows the value of open-weight AI. Developers can adapt existing models and turn them into something completely new, without starting from scratch. But it also demonstrates the need for strong open-weight models from democracies. Developers should not be forced to only build on top of systems that may reproduce Chinese Communist Party propaganda.
Thinking Machines Announces Credits for Safety Research
Thinking Machines, a US AI company that builds open-weight models, launched a new grant program offering researchers up to $50,000 in AI credits to study how open-weight AI models can become safer and more resilient to misuse. Since open-weight models can be modified, attackers can remove built-in safeguards and use the models to automate cyberattacks. But this same capability means that researchers and developers can also modify models to build and patch vulnerabilities in important technology. These grants will fund research into making open-weight models more useful for defenders than attackers, while also addressing broader safety questions to minimize potential harm from open-weight AI.
Prem Announces an AI Agent that Continuously Scans Code
Prem, a private AI infrastructure company, launched Cyberscan, an AI agent that works 24/7 to detect vulnerabilities in codebases, particularly those around the censorship-resistant currency Bitcoin. The launch follows a wave of AI-driven attacks on Bitcoin projects, including an exploit of Coldcard hardware wallets that has drained more than $116 million since July 30. In response, Prem and software developers began using AI to scan Bitcoin-related codebases for weaknesses that attackers could exploit. Now, Cyberscan makes this assistance more automatic. Developers can connect their codebases (Bitcoin related or not), and the agent continuously scans the code, looks for possible vulnerabilities, and flags anything suspicious. CyberScan exclusively uses open-weight models, as they give Prem more control over how the models operate and reduce reliance on the restrictions of closed AI providers. Prem says the tool has already helped over 20 teams find critical security issues.
Why this matters: Bitcoin offers permissionless, censorship-resistant money to people living under surveillance, inflation, and financial repression. But that promise depends on the security of the open-source infrastructure around it. Cyberscan helps catch vulnerabilities faster and keeps Bitcoin safer for the people who rely on it most.
Liquid AI Makes Local Models Smaller and Faster
Liquid AI — a US AI company that focuses on making small, open-weight models — has released two updates that make its models even easier to run on personal devices. First, it released compressed versions of several of its models. That means that they modified the models so that they require less memory (using fewer of a device’s resources while running) while keeping nearly all of their original performance. The next day, it released another update that allows some of their models to generate responses over twice as fast, all without reducing answer quality. Together, the updates make Liquid AI’s models lighter, faster, and more practical to run locally on devices like laptops and phones.
Recommended Content
LOCAL LOCAL LOCAL (AI) with Mark Suman
Bitcoin Park Co-Founder’s Rod Roudi launched a new show, Tangents, in which he explores updates in freedom technology. In a recent edition, he sat with Maple AI CEO Mark Suman for a 10-minute conversation on the current landscape of AI and the future of privacy. Suman discusses why “not all privacy is created equal,” explains the difference between AI proxies and end-to-end encrypted AI options like Maple, and considers what recent AI news could mean for user privacy and control. Give it a listen here.
