HRF's AI for Individual Rights Newsletter #23

• THE LATEST IN AI FOR REPRESSION
• RECOMMENDED CONTENT
• THE LATEST IN AI FOR FREEDOM
• RECOMMENDED CONTENT

Welcome to the 23rd edition of HRF’s AI for Individual Rights newsletter.

Uganda’s regime, led by dictator Yoweri Museveni for over four decades, has long suppressed political opposition — and its repression is now supercharged by foreign surveillance tech. Chinese facial recognition cameras monitor the streets, Russian tracking technology has been used to follow vehicles, and forensic tools extract information from dissidents’ phones. These technologies make it increasingly difficult for dissidents and political opponents to organize and remain safe.

Meanwhile, US tech giant Meta has released what it calls the “most capable” American open-weight model yet. Open-weight models can be downloaded, modified, and run privately, and therefore give people more control over their AI use, without depending on corporate access. With China currently leading in open-weight AI, Meta’s release is an important step toward making powerful AI developed in a democracy more widely accessible.

We close with a fireside chat with HRF’s Alex Gladstein at Bitcoin Policy Institute’s Freedom Tech DC last month. He discusses the rapid evolution of AI, HRF’s work on private AI agents and organizational intelligence, and how these technologies could shift the balance of power toward individuals advancing freedom.

§THE LATEST IN AI FOR REPRESSION

§Uganda Demonstrates the Brutality of AI-Enabled Authoritarians 

In a New York Times guest essay, reporter Megan Stack describes a “reign of terror” in Uganda, now assisted by foreign surveillance technology. Chinese facial recognition cameras monitor Uganda’s streets. Russian-supplied technology has been used to track vehicles through chips embedded in license plates. Ugandan police have even acknowledged using these tools to track supporters of opposition leader Bobi Wine. One Ugandan novelist, Kakwenza Rukirabashaija, was repeatedly detained and tortured for criticizing the regime. He later discovered that officials had extracted data from his phone, analyzed his social media posts using Israeli Cellebrite software, and pulled his cloud data with Canadian Magnet Forensics tools. Although the regime imports this technology from around the world, Stack particularly warns readers about imports of China’s surveillance technology, as well as its authoritarian approach to governance, spreading to Uganda.

In context: Uganda’s 82-year-old president, Yoweri Museveni, has ruled since 1986 and subjected the population to decades of censorship and violent repression of dissent. His son, military chief Muhoozi Kainerugaba, has more recently been accused of overseeing the detention and torture of political opponents. In January 2026, Bobi Wine challenged Museveni, resulting in a violent crackdown on Wine’s supporters. The expansion of surveillance technology has coincided with this growing opposition, and could make it increasingly difficult for dissidents to survive.

§Hong Kong Police Launch AI Platform to Teach Children National Security

Hong Kong police have launched the NSmarties AI Interactive Platform, a state-led educational tool that shares official narratives about national security directly with primary school students. With generative AI, children can create personalized stories on China’s “holistic approach to national security,” a doctrine promoted by Xi Jinping that places protecting the Communist Party’s rule at the heart of national security. Students can also complete interactive quizzes that may reinforce repressive state narratives and listen to explanations of security laws narrated by Hong Kong’s police commissioner and security chief.

Why this matters: China imposed its draconian National Security Law in 2020. Since then, officials have used broadly defined national security offenses to imprison pro-democracy activists, shut down independent media, and suppress political opposition. AI could help normalize laws that enable political repression and discourage future dissent.

§AI Coding Tool OpenCode Hit With a Security Breach 

OpenCode, an open-source AI coding agent, disclosed a security breach affecting some users of its online platform, OpenCode Console. These users had connected their OpenCode agents to external AI providers (such as OpenAI, Anthropic, and Google) using API keys (credentials that charge AI usage to users’ accounts). Attackers gained access to an older system storing these API keys. The keys appear to have been stolen and used to access AI services at users’ expense, leading to unauthorized charges on some accounts. The company has not confirmed who was behind the attack but claims it found no evidence that other user data was exposed. The breach is a reminder that every third-party AI service that stores sensitive credentials can be breached. Running an OpenCode agent locally (by powering it with local models instead of third-party ones like OpenAI) mitigates this risk and gives users greater control over their data and security. 

In context: OpenCode tried using closed-source AI to defend against the attack. Sometimes those models refuse to help as their safety guardrails block them from analyzing the attack. With closed-source AI, the provider can deny access even to those who need it in critical moments. Luckily, the team later gained access to OpenAI’s Daybreak Blue program, which lets approved cybersecurity professionals use AI models with fewer restrictions. 

§Targeted Ads Turned Into a Zero-Click Spyware Pipeline

The Citizen Lab documented the first publicly confirmed case of spyware delivered through targeted online advertisements that could infect devices without users even clicking on them. Israeli company Discover-Tech purchased and placed ads that delivered spyware developed by Candiru, another Israeli company. Candiru’s spyware can steal private information, is sold exclusively to governments, and has been used to target dissidents and journalists. Citizen Lab researchers confirmed that at least 45 devices at the news organization Al Jazeera were targeted in 2024 and 2025 using this digital ad technique, although they could not identify who was behind the attacks. The ads also secretly directed millions of devices to Candiru’s servers, likely to track users and identify potential targets. This is a new form of surveillance, one that is hard for an individual to detect and could be sold to authoritarian regimes to secretly target and spy on dissidents.

Why this matters: The Citizen Lab did not report that this was an AI-powered attack. However, as AI makes it faster and cheaper to develop software, similar surveillance techniques could become easier to develop and deploy.

§China Proposes Export Controls on AI Models

China has proposed new export restrictions on 23 AI technologies, including AI models. The rules would require approval before Chinese companies could transfer certain powerful AI models or tools overseas. The proposal is still under review, but China has already begun requiring companies to register exports of the technologies covered, including certain advanced AI models. A commerce ministry spokesperson said the rules aim to “prevent misuse and abuse” and “ensure that AI remains under human control.” The proposed rules could slow down or even prevent Chinese companies from openly releasing advanced models, the very ones that could empower individuals.

Why this matters: China is a leader in developing open-weight AI, models that anyone with sufficient hardware can download, modify, and run themselves. This gives dissidents the freedom to use the same AI models available to the regimes that try to repress them, evening the playing field. The weights already released would stay legal to share. But these rules risk deterring Chinese companies from releasing the weights of future frontier models. That could concentrate control over tomorrow’s most powerful AI in the hands of corporations (and the Chinese Communist Party that can exert control over them), rather than the individuals who need it most.

§RECOMMENDED CONTENT

§How Generative AI Is Scaling Transnational Repression

In the independent policy journal Just Security, three researchers from The Citizen Lab analyze findings from Anthropic’s September threat intelligence report to illuminate how authoritarian actors are using AI to surveil and target dissidents abroad. For example, a China-aligned operator used Claude, Anthropic’s AI assistant, to track Uyghurs living in Syria and recruit informants within the community. Read the article here to discover how generative AI helps intimidate dissidents across borders.

§THE LATEST IN AI FOR FREEDOM

§Meta Releases Open-Weight Model “Muse Spark Open” 

Meta has released Muse Spark Open, what the company calls “the most capable US open-weight model to date.” The tech giant says the model excels at coding and AI-agent tasks. But Muse Spark Open still requires extremely expensive hardware to run. One day it could be hosted on privacy-preserving services that can host open-weight models at much lower cost. And because its weights are open, any developer could compress the model so it can run on less expensive devices. In fact, Meta worked with Apple to make the model run locally on two M5 Ultra Mac Studios. This could make it a potential option for nonprofits seeking to operate their own AI infrastructure.

In context: Chinese AI companies currently dominate the open-weight AI landscape, and many of these models have been found to reproduce Chinese Communist Party narratives. Muse Spark Open marks a step toward interrupting this dominance. 

§NVIDIA Introduces Three New Open-Weight AI Models

NVIDIA has introduced three new open-weight models designed to power AI agents. The largest of the models, Nemotron 3.5 Super VL, is built for complex tasks that can span days or weeks, but requires extremely powerful and expensive hardware to run locally. Its other release, Bolt 74B, can power AI agents that run locally on high-end computers like NVIDIA’s DGX Spark. Both models are available to download, and NVIDIA plans to release the smaller Bolt E4B on Oct. 23, designed to bring capable AI agents to more accessible, everyday devices like consumer laptops. These models could allow individuals and organizations to build customized AI agents that run privately on their own hardware, without relying on corporate cloud services.

§Release of DeepSeek-V4.1-Flash-1017 Brings Praise from Local AI Community

Chinese AI company DeepSeek has released DeepSeek-V4.1-Flash-1017, a capable new open-weight AI model. According to DeepSeek’s benchmarks, the model outperforms Anthropic’s Claude Opus 5 and OpenAI’s GPT-5.6 Sol on several coding and AI agent tests, while being significantly cheaper. It has also received significant praise from the local AI community, including Osmantic founder and CEO Ahmad Osman, who said the model is “genuinely the most impressive … open source model I’ve ever worked with.” Local AI developers are already compressing it to run on less expensive hardware so more people can interact with the model privately. It is also now available on Maple and Tinfoil, two privacy-focused AI services that claim to protect users’ conversations with end-to-end encryption. It’s exciting to see how quickly independent developers are finding ways for individuals to benefit from powerful Chinese AI models without sending their private conversations to servers in China.

Important: Research has found that previous DeepSeek models can reproduce Chinese Communist Party narratives, particularly on topics such as government accountability, political dissent, and civic activism. Be aware of these potential biases and carefully evaluate the model’s responses. 

§Privacy Browser Brave Adds Encrypted AI Chat 

Privacy-focused browser Brave has introduced end-to-end encryption for its built-in AI assistant, Leo. Users can ask Leo questions, summarize webpages and videos, translate content, and more. And now, they can do so more privately. Users’ questions to Leo are encrypted before leaving their devices and processed inside Trusted Execution Environments (secure hardware environments that prevent Brave and the AI provider from seeing the user’s AI interactions). The assistant currently uses the Chinese open-weight model GLM 5.3 Flash. Brave was already a powerful tool for private browsing and searching. Now, with built-in private AI, it’s making the internet even more useful without compromising users’ privacy.

§Indian Developer Turns Protest Repression Into a Video Game

Indian software engineer Suraj Kushwah created a satirical new video game, “Mantar Jantar,” that lets players experience the challenges of protesting in New Delhi. Players must navigate police barricades, evade facial recognition technology, avoid tear gas and arrest, and reach Jantar Mantar, the city’s historic protest site. Kushwah has found a creative way to expose the repression protesters are currently facing.

In context: India has recently seen a wave of youth-led protests over unemployment, examination scandals, and alleged electoral irregularities. Recent demonstrations have been met with police violence, internet shutdowns, and the detention of at least 10,000 people, according to Amnesty International. The Delhi Police also deployed AI-powered facial recognition at Jantar Mantar earlier this year. Kushwah’s game brings these tactics to life, exposing the obstacles to peaceful protest.

§RECOMMENDED CONTENT

§How Private AI Helps Human Rights Activists

At the Bitcoin Policy Institute’s (BPI’s) Freedom Tech DC last month, HRF’s Alex Gladstein sat down with BPI Senior Director of Education Andrew Bailey to discuss the rapid evolution of AI and its potential to empower activists. Gladstein shared HRF’s work with AI startup Finite to build privacy-protecting AI agents and organizational brains, explained the importance of open-weight models for privacy, and made the case for democratic leadership in open-weight AI. Watch it here.