HRF's AI for Individual Rights Newsletter #24

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

Welcome to the 24th edition of HRF’s AI for Individual Rights newsletter.

In Vietnam, the Communist Party is ramping up its use of artificial intelligence (AI) to censor online criticism and intensify surveillance. It is also expanding VNeID, a digital identity app used by more than 67 million citizens, into an “AI-powered surveillance super-app.” In The Vietnamese Magazine, writer Aerolyne Reed makes the case for pro-democracy activists embracing AI in response, and even draws on HRF’s AI for Activists guide.

National Public Radio (NPR) also highlighted AI’s potential to empower human rights defenders on “All Things Considered.” The segment featured HRF’s AI Hack for Freedom III, held Sept. 24–25 in Washington, D.C., where human rights defenders teamed up with developers to turn their real-world human rights challenges into working solutions.

In a recent issue of China Journal of Democracy, HRF’s Alex Gladstein echoes that narrative and explains that AI agents and privacy-protecting AI can multiply the work of underfunded Chinese human rights organizations. He also explores the irony of Chinese AI models helping dissidents challenge the Chinese regime. These models can be accessed in privacy-protecting ways, giving users powerful intelligence without sending data to Chinese servers or even off the user’s computer.

§THE LATEST IN AI FOR REPRESSION

§Vietnam’s AI-Enabled Repression and How Dissidents Can Resist 

In a recent article for The Vietnamese Magazine, writer Aerolyne Reed examines how Vietnam’s Communist Party uses AI to intensify its censorship and surveillance of civil society and dissent. In December 2025, dissidents Lê Trung Khoa and Nguyễn Văn Đài were each sentenced to 17 years in prison over accusations of using AI-generated deepfakes to mock party leaders. More recently, leaked documents revealed that the party wants 5,000 AI specialists and 1,000 social media influencers to enforce “ideological immunity,” and hopes that AI can remove 90% of prohibited online content within 24 hours. Vietnam also plans to expand its digital identity app, VNeID, used by more than 67 million citizens, into “an AI-powered surveillance super-app.” It would feed citizens’ biometric data, financial records, private communications, and public CCTV footage into AI models for the state to analyze. It’s becoming even harder for dissidents to speak and work freely without the state’s AI knowing. 

Why this matters: Reed argues that “pro-democracy advocates in Việt Nam must utilize AI to have any chance of success.” She cites HRF’s AI for Activists Guide to show how it can transform the work of those with limited resources. To evade state surveillance, an activist can run AI models locally so no data ever leaves their device, or use encrypted services. Personal AI agents can help with research, monitor state media, and alert trusted networks if an activist is detained. 

§Belarus’ National AI Framework Could Strengthen Surveillance 

In a recent TV interview, Belarus’ communications and information minister, Kiryl Zalesski, revealed that the regime is building a “national artificial intelligence framework,” one that could expand its surveillance capabilities. Zalesski explained that 60,000 surveillance cameras already feed into AI to automatically detect traffic violations. But in a regime that has used video footage to identify and arrest anti-regime protesters, these capabilities could quickly be turned against dissidents. Zalesski also announced plans to adapt open-weight AI models (models that anyone with sufficient hardware can download and run themselves) and host them on domestic servers. This could give the regime greater capability to use AI to build surveillance tools, track political opponents, and analyze citizens without relying on foreign technology providers or their oversight.

Why this matters: Open-weight AI models allow anyone to adapt powerful models for their own purposes. While this can empower an individual dissident, it can also give authoritarian regimes greater independence to enable AI-powered repression.

§France Deepens AI Partnership With the UAE Despite Surveillance Concerns

French President Emmanuel Macron and Abu Dhabi Crown Prince Khaled bin Mohamed bin Zayed Al Nahyan attended the United Arab Emirates (UAE)–France Business Dialogue, where the countries deepened their AI cooperation. French AI company AMI Labs signed agreements with Emirati organizations to explore collaboration on models, cloud computing, research, and robotics. Among them is G42, a UAE-based AI company whose executives previously held senior positions at a surveillance firm involved in hacking journalists, activists, and political opponents on behalf of the regime. The partnership raises concerns that French AI expertise could strengthen the technological capabilities of a regime with a history of using surveillance to repress dissent.

§Hong Kong Police Expand Surveillance Network to Shopping Malls

Hong Kong police are planning to connect cameras in private shopping malls to SmartView, the city’s expanding AI-assisted surveillance network. Security chief Chris Tang said police could then access mall footage, and added that real-time monitoring of certain locations could be necessary. The expansion is part of a broader plan to install at least 60,000 cameras by 2028 and introduce AI-powered facial recognition. While officials cite public safety, the growing network raises concerns about the regime’s ability to identify and track dissidents across the city.

§Chinese State-Linked Hackers Use AI Agents to Attack Hundreds of Organizations

Microsoft has revealed that a Chinese state-linked hacking group, Storm-0940, used AI agents to automate cyberattacks against hundreds of unnamed organizations worldwide, including governments, media outlets, think tanks, and nonprofits. Using ARTEX, an open-source AI hacking framework, the group automated the process of finding security weaknesses, attempting break-ins, and gaining access to internal systems. Most of the login attempts were unsuccessful, but in several cases, attackers gained high-level access to organizations’ cloud infrastructure in less than three hours. Microsoft says attackers wanted to steal data, but did not reveal the broader motive. 

Why this matters: The inclusion of nonprofits and think tanks in such a widespread cyberattack is particularly concerning. While we don’t know whether any of the targeted organizations were critical of the Chinese regime or involved in human rights work, the campaign demonstrates how easily state-backed hackers could use AI to target organizations resisting authoritarian rule.

§RECOMMENDED CONTENT

§Human Rights Activists Look to Harness AI for the Greater Good

HRF’s AI Hack for Freedom III was featured on NPR’s “All Things Considered” podcast. The segment explores the impact of our two-day hackathon, held Sept. 24 and 25 in Washington, D.C., which brought together human rights defenders and developers to use AI to build tools for freedom. NPR cybersecurity correspondent Jenna McLaughlin spoke with participants about their experiences putting AI to work for human rights. Evan Mawarire, a Zimbabwean pro-democracy activist who led the ThisFlag movement, explained, “If activism is going to be effective, it also has to evolve.” AI is opening up new possibilities for activists to do just that. Listen to the segment here.

§THE LATEST IN AI FOR FREEDOM

§Recursive Releases Powerful AI Coding Model That Runs on a Laptop

AI startup Recursive has released Flint, an AI model that founder and CEO Richard Socher called the “best open-weight model for running a coding agent on your laptop.” The model is designed to write code and fix software bugs directly on accessible consumer devices. According to Recursive’s benchmarks, Flint even outperforms similarly sized models on several coding tests, while running up to twice as fast. With Flint, human rights activists and organizations could gain the ability to build tools and develop new software they need without sending any confidential information off their devices. 

In context: Recursive says Flint was built from scratch by a small team of people and their own automated AI research system in under four months. That means AI helped build AI, demonstrating that small teams and startups without massive amounts of funding can quickly build powerful tools that could accelerate the benefits for individuals. 

§Underdog Launched Local Speech-to-Text Model Bark 3.0

American AI startup Underdog has released Bark 3.0, a speech-to-text open-weight model that is small enough to run on a laptop or smartphone. The company says it is as accurate as NVIDIA’s Parakeet, a transcription model 15 times bigger, and even beats OpenAI’s Whisper large-v3, which is 57 times bigger. Anyone can now download the model and run it to dictate a press release, turn a brainstorm session into searchable text, or generate notes from an important meeting, all privately on their phone. 

In context: Underdog is also developing a local AI assistant that handles chat, email, calendars, video generation, and more, all powered by open-weight AI models. Bark 3.0 now accelerates Underdog’s transcription abilities, but the agent’s broader software is closed source, so its privacy and security claims cannot be verified independently. 

§Cohere Adds Confidential Computing for Encrypted AI Prompts

Cohere, a Canadian AI company, has announced a new way to interact with its models privately. Its online platform, Model Vault, can now encrypt users’ prompts, documents, and responses before they leave their devices and process them inside secure hardware environments, called Trusted Execution Environments (TEEs). These protections are designed to prevent even Cohere and its cloud provider from accessing the contents of users’ AI interactions. Cohere designed this feature for enterprises, but there is potential for human rights organizations and individuals to use it to safely analyze confidential documents and other sensitive information without needing to operate their own expensive AI infrastructure.

§Block Releases Dessa, A Speech Model that Runs Locally

Block, the payments company created by Jack Dorsey, launched Dessa, a speech-to-speech model that the company says feels like talking to “a person … full of emotion, nuance, pauses, and a natural cadence.” Dessa can run locally on a high-end consumer laptop, such as a MacBook Pro M4 Max. This could eventually allow human rights defenders to interact naturally with AI without sending their conversations to the cloud. For now, however, Dessa’s model weights are not publicly available, and access is limited to select AI safety researchers.

Why this matters: Dorsey said, “We believe voice is the most natural way to interact with intelligence.” If this is true, more and more people will want to literally speak with AI, and it is important that dissidents have a local, private way to do so.

§RECOMMENDED CONTENT

§HRF’s Alex Gladstein in China Journal of Democracy 

HRF’s Chief Strategy Officer Alex Gladstein was featured in China Journal of Democracy alongside four experts discussing how AI could empower democratic movements. He explains how AI agents and privacy-protecting tools are force multipliers, especially for Chinese human rights organizations operating in exile, many of which depend on just one or two grants. He also explores a striking paradox: Chinese companies are releasing some of the world’s most powerful open-weight AI models, which can be downloaded and used privately, without sending data to Chinese servers. These models can run in TEEs that protect conversations from providers, while smaller versions can run on personal devices. This could give dissidents greater control over powerful AI tools, even inside China under the regime’s extensive surveillance, though whether officials can effectively restrict local AI use remains an open question. Read it here.