HRF's AI for Individual Rights Newsletter #26

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

Welcome to the 26th edition of HRF’s AI for Individual Rights newsletter.

Hong Kong police plan to integrate AI facial recognition into the city’s network of surveillance cameras this month. In a city where cameras have already been used to monitor pro-democracy protests, an AI-powered ability to automatically identify and track dissidents could take repression to new heights.

Meanwhile, privacy-focused company Maple AI has launched a powerful, privacy-protecting AI agent called Maple Work. It can do real work on computers, edit and organize files, and automate complex tasks, without handing data to the AI provider or Maple itself. At the same time, American startup Thinking Machines Lab released an open-weight model that potentially challenges China’s dominance in open AI.

We also include an episode of HRF’s podcast, “The Dissident’s Mic,” in which one of Afghanistan’s first female tech CEOs, Roya Mahboob, shares how she is using technology, from Bitcoin to offline AI, to uplift thousands of girls living under Taliban rule.

§THE LATEST IN AI FOR REPRESSION

§Hong Kong Police to Add AI-Powered Facial Recognition to Its Surveillance Network

This month, the Hong Kong Police announced plans to integrate facial recognition capabilities into SmartView, the city’s camera surveillance network. With facial recognition capabilities, police aim to automatically search through camera footage for suspects and receive real-time alerts when certain individuals are detected. Facial recognition data will be kept for around 180 days, six times longer than ordinary CCTV footage, raising concerns about data privacy. The rollout is part of a broader plan to grow the system from roughly 10,000 cameras to 60,000 by 2028, all in the name of public safety and crime prevention.

Why this matters: Hong Kong police have previously used video cameras to monitor pro-democracy protests and collect footage for investigations. Integrating facial recognition into the city’s surveillance network could automatically track the physical location of people the regime deems undesirable and expand the regime’s capacity for repression.

§Qatar’s AI Market Could Be Worth More Than $2 Billion

A new report by Strategy&, part of consulting firm PwC, says Qatar has an opportunity to tap into a domestic AI market worth more than $2 billion by 2030 and export its technology across the region. The report encourages Qatar to use state agencies, hospitals, banks, and other major institutions to develop and test its AI applications. It also highlights the role of state-backed AI company Qai in building Qatar’s own AI infrastructure and computing capabilities. While this could be a great economic opportunity, the strategy could also help bring stronger AI capabilities to a regime that heavily restricts political dissent. And as Qatar builds out these AI systems, the regime could export them into the hands of other authoritarians.

In context: Qatar is ruled by a fully authoritarian regime that has imprisoned peaceful critics, including activists challenging discriminatory election laws and advocates for migrant workers. Its cybercrime law criminalizes spreading vaguely defined “false news” and gives officials broad powers to restrict online expression. A $2 billion AI market could hand the regime greater power to monitor critics and online speech, and to repress its population.

§China’s Plan to Regulate AI Industry

China’s National Development and Reform Commission said it would accelerate comprehensive legislation to regulate how Chinese AI is developed and used, covering everything from model development and training data to computing infrastructure. The new rules would reportedly set safety standards (whose details have not yet been specified and could be subject to change) and increase state oversight of AI companies such as DeepSeek, Moonshot, and Alibaba. These companies are making some of the world’s most powerful open-weight models, ones that can be downloaded, modified, and run privately to give individuals more control over their AI use. Beijing frames the regulations as necessary for AI safety and promises to keep supporting open AI development. But heavier regulation could discourage companies from openly releasing future models, thereby limiting access to technologies that empower individuals.

Why this matters: The CCP already exercises control over Chinese AI companies, requiring models to uphold the regime’s ideals. Greater regime oversight could ensure these CCP narratives remain ingrained in AI while giving officials more influence over how the companies handle user data, potentially enabling greater surveillance of citizens and dissidents.

§Apple Warns Users in 99 Countries of Sophisticated Spyware Attacks

Apple has reportedly sent warnings to users in 99 countries saying they may have been targeted by sophisticated mercenary spyware, according to threat notifications shared by recipients. Among those receiving alerts were Ragip Soylu, a Turkish journalist, and Radwan Mortada, a journalist and filmmaker who was sentenced to prison in 2021 for allegedly criticizing the Lebanese military. Apple has not identified who is behind the latest attacks or confirmed that the targeted devices were successfully compromised. These warnings highlight the global reach of spyware that can compromise the devices of activists and human rights defenders and expose their private information and networks.

Why this matters: AI could allow attackers to develop and deploy spyware at an increasing rate. Dissidents in authoritarian states should follow Apple’s guidance for protecting against mercenary spyware and keep their devices updated.

§RECOMMENDED CONTENT

§Can Technology Open Doors for Women in Afghanistan?

In a recent episode of HRF’s podcast, “The Dissident’s Mic,” Roya Mahboob — one of Afghanistan’s first female tech CEOs, now living in exile — reflects on five years of Taliban rule and her work using new technologies to empower Afghan women and girls. She shares how she founded the Afghan Dreamers, an all-girls robotics team, and how she uses the permissionless digital currency Bitcoin to pay women facing barriers to traditional banking. With HRF’s support, Mahboob is developing an offline, AI-powered tool to help Afghan girls continue learning despite the Taliban’s ban on secondary and higher education.

§THE LATEST IN AI FOR FREEDOM

§Maple Launches Private AI Agent That Can Work Across Your Computer

Maple AI has launched a privacy-protecting AI agent called Maple Work. Unlike conventional agents, Maple Work is open-source and powered by open-weight AI models running in Trusted Execution Environments (TEEs). These are secure hardware environments that hide users’ AI interactions from both Maple and the companies hosting the models. Users can run and steer tasks on their other computers over an end-to-end encrypted link. Agents can also search the web, easily connect to 26 external apps, and even hand off tasks to more powerful agents like Claude Code or Codex when needed, though information shared with external services is not covered by Maple’s enclave privacy protections. Maple Work is a breakthrough invention that is especially exciting for human rights defenders. Imagine the possibilities of a simple-to-use, powerful AI agent that automatically researches and drafts funding applications, monitors state media, or investigates corruption, all without exposing sensitive prompts to AI providers.

In context: When powering agents through mainstream AI providers, users must simply trust that the providers are running the model they claim to be. But Maple Work, because it uses TEEs, allows users to cryptographically verify which model is powering their agent. This is especially valuable for activists, who could otherwise unknowingly be served a less private or deliberately censored model.

§Thinking Machines Releases Inkling 2 Small That Challenges Chinese Dominance

American AI startup Thinking Machines Lab has made an exciting contribution to the open AI ecosystem with the release of Inkling 2 Small, an open-weight model designed for coding, research, and complex reasoning. Artificial Analysis, a leading independent AI benchmarking platform, has ranked it as the most intelligent open-weight model developed outside of China. Inkling 2 Small requires expensive hardware to run locally, but developers could optimize and compress it for more accessible devices or run it in secure environments, enabling civil society to interact with it in a privacy-protecting manner.

In context: AI models developed in China, which have been found to reproduce CCP narratives, dominate the open-weight AI intelligence rankings, according to Artificial Analysis. Inkling 2 Small is disrupting this pattern and now lets dissidents access powerful and private AI that was developed in a democracy.

§Nous Research Post-Trains Chinese AI Model to Reduce CCP Bias

Nous Research, the company behind the open-source agent Hermes, released Arete-27B, a model built to strip the CCP’s propaganda out of one of China’s most popular open models, Qwen3.8-27B. Researchers post-trained it using answers generated by another open-weight model, Inkling. The goal was to make Qwen respond with less political bias to sensitive questions about democracy, freedom of expression, and surveillance. In Nous Research’s tests across 803 questions, the proportion of responses containing Chinese state-aligned narratives dropped from 22% to 0.9%. Arete-27B can run locally on sufficiently powerful machines (like an NVIDIA DGX Spark), expanding access to powerful, less-censored AI without sending sensitive conversations to external servers.

Why this matters: Even open-weight AI developed in an authoritarian regime can be transformed to challenge state censorship and bias and offer a powerful and private tool for dissidents to access information.

§Amazon Launches Nitro Sandboxes for More Secure and Private AI Agents

Amazon Web Services has launched Nitro Sandboxes, a new feature that allows AI agents to run in isolated, hardware-protected environments. Sandboxes restrict what an AI agent can access or do. For example, a sandboxed agent could help a journalist analyze leaked documents on corruption without being able to access the journalist’s contacts or send anything out to the web. Nitro Sandboxes goes beyond providing a sandbox and also offers a confidential computing mode. This feature uses hardware protections to keep the information an AI agent processes private and hidden from the administrators operating the servers and from the companies hosting the AI. Developers could use Nitro Sandboxes to make more private and secure AI agents for human rights defenders.

§Google Introduces New AI System Designed for Private AI

Google has announced a new AI computing system, TPU 8tx, designed to make AI more secure and efficient. Tensor Processing Units (TPUs) are Google’s specialized chips for training and running powerful AI models. TPU 8tx introduces hardware for confidential computing, which allows AI models to run on Google’s chips while keeping the user’s prompts and data hidden. This could allow developers to build powerful, privacy-protecting AI services that help civil society leaders and human rights activists accelerate their work without exposing information to the companies hosting the AI.

§Underdog Releases Text-to-Speech Model Capable of Running on a Phone

Underdog, an AI startup building local AI tools, has released Woof, a local text-to-speech model that reads text aloud in a natural voice. Woof comes in two versions: a smaller, single-voice model and a larger, two-voice model. The one-voice model can run on a phone or directly in a web browser, even offline. And the two-voice Woof beats Fish Audio Drama 3 — a voice model over 1,000 times its size — reading numbers, dates, prices, and codes with 25% fewer errors, according to Underdog. Dissidents could use it to read private messages and documents without sending a word off their device.

§RECOMMENDED CONTENT

§Maple AI’s Privacy Timeline

Maple AI, an AI company providing encrypted AI inference, has released a privacy timeline that tracks changes in AI companies’ privacy policies, data breaches, government surveillance of AI, and increases in regulations. Since 2022, it has covered 178 developments involving major AI providers, including OpenAI, Google, Anthropic, Microsoft, and xAI. To stay informed about where your data is going when you use AI tools and how privacy protections could change over time, check it out here.