HRF’s AI for Individual Rights Newsletter #14

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

Welcome to the 14th edition of HRF’s AI for Individual Rights newsletter.

We begin in Afghanistan, where the Taliban is using tens of thousands of cameras to surveil women forced to follow its repressive rules. On a similar note, Indian police are planning a hackathon to build AI tools that can search surveillance footage and identify people of interest, potentially making it far easier to track protesters, dissidents, and ordinary citizens at scale.

At the same time, new sovereign tools are also empowering the individual. Chinese tech company Alibaba released Qwen 3.8 27B, an open-weight AI model. That means individuals with the right hardware can download, modify, and run this file themselves. This model is small enough to run on high-end computers, but still offers powerful intelligence. It is an encouraging step toward putting more advanced AI directly in the user’s hand, and a path that AI companies from democratic societies are beginning to follow.

We end with a report from Hugging Face — the leading platform for downloading open-weight AI models — that examines how the open-weight landscape has grown and changed this year.

SUBSCRIBE TO THE NEWSLETTER HERE

§THE LATEST IN AI FOR REPRESSION

§Kabul’s 90,000 Cameras Put Afghan Women Under Constant Watch

Women in Afghanistan report that constant camera surveillance has made life under Taliban rule feel like a prison. Since its return to power five years ago, the Taliban has installed over 90,000 Chinese-built security cameras in the capital city, Kabul. At the same time, it has issued over 100 decrees restricting women’s participation in public life. Officials can now use the cameras to identify and punish women for an improperly worn headscarf, speaking in ways deemed inappropriate, or being outside without a male guardian. Tahmina, a former teacher living in Kabul, described her fears: “You feel you are constantly being watched. They are waiting for you to break their laws so they can come and take you away.“

Why this matters: Officials claim the security cameras have AI facial recognition. If that is true, or if any other AI-powered analytics are integrated into the surveillance system, the Taliban could enforce its extreme restrictions on women with far greater reach and precision.

§Police in India to Host a Hackathon to Connect Surveillance Cameras

Police in the Indian state of Gujarat announced plans to host a major hackathon aimed at connecting Gujarat’s 80,000 police cameras into a single surveillance network. The cameras currently run across different government departments and technological systems, which makes it difficult for police to monitor people and incidents between cameras. To make this surveillance more seamless, the hackathon will gather students, researchers, and technology companies to build AI tools that can scan live feeds, recognize suspicious activity, match people against watchlists, and alert the police. If successful, the event could transform a network of unconnected cameras into a system for a hybrid authoritarian regime to watch protesters, journalists, and other critics.

In context: Just last month, officials in India deployed AI-powered facial recognition technology to track student protesters. If that pattern continues, future protesters could face an even more powerful surveillance system.

§North Korean State-Linked Hacking Group Integrates AI for Cyberattacks

Kimsuky, a cyber-espionage group controlled by the North Korean regime, is reportedly integrating AI to scale its cyberattacks. The group has a history of using phishing campaigns to steal information and gather intelligence on foreign policy, sanctions, and other issues relevant to Pyongyang’s strategic objectives. Now, researchers say, Kimsuky is building the capacity to use privately run AI models for malware development, data analysis, and the automation of cyberattacks. Kimsuky could then more rapidly steal and analyze sensitive information about governments, researchers, journalists, and even organizations working to advance human rights in North Korea. In the hands of one of the world’s most repressive dictatorships, that capability could make espionage faster, broader, and far more damaging.

§Moscow Police Beat Innocent 15-Year-Old After AI Facial Recognition Misidentifies Him

Two Moscow police officers beat a 15-year-old schoolboy after an AI-powered facial recognition system falsely flagged him as an 85–90% match with a suspected criminal. The officers reportedly attacked him from behind, knocked him to the ground, left him with a concussion and bruises, and accused him of distributing drugs. AI-powered surveillance need not be accurate to be dangerous in the wrong hands. In repressive systems, a false match is enough to become grounds for detention, abuse, or worse, giving officials a technological pretext to target innocent people while obscuring human responsibility behind an algorithm.

Why this matters: This story also serves as evidence that Russian police are actively using their AI-powered surveillance system. For dissidents and protesters, that means attending a demonstration — or simply moving through Moscow — could make them easier for authorities to track and target.

§Apple Builds China-Specific AI Model Under CCP Rules

Apple has reportedly trained an AI model specifically for users in China, potentially adapting it to the country’s censorship rules. Developed in partnership with Chinese tech giant Alibaba, the model is intended to power Apple Intelligence (Apple’s on-device AI) on iPhones and Macs sold in China. Apple could not simply deploy its usual AI system because generative AI services operating in China are subject to strict controls. AI providers must uphold the Chinese Communist Party’s (CCP’s) “core socialist values,” suppress prohibited content, and comply with regime censorship requirements. That means Apple may have trained an AI that echoes CCP-approved narratives and hides information about the regime’s human rights abuses.

In context: This would not be the first time Apple has complied with CCP demands. The company has previously removed censorship circumvention and messaging apps from its App Store in China. Authoritarian regimes can pressure technology companies seeking access to their markets to restrict what users can access. And now, that censorship may be extending into AI.

§RECOMMENDED CONTENT

§Teaching AI as a Tool for Freedom

In a recent edition of the HRF x PubKey Freedom Tech Series, Sarah Moorman, HRF’s Freedom Tech content associate, sat down with Anna Chekhovich, HRF’s Freedom Tech education lead, and Alex Lewin, lead engineer at Finite. They discussed how dictators use AI to repress and, more encouragingly, how AI and privacy-protecting agents can supercharge activists on the frontlines of the struggle for human rights and democracy.

§THE LATEST IN AI FOR FREEDOM

§New Qwen Model Brings Frontier-Level AI Closer to Personal Devices

Chinese tech giant Alibaba released a small but mighty open-weight model, Qwen 3.8 27B. Open-weight models are AI systems that individuals with sufficient hardware can download, modify, and run themselves to interact privately with AI. Often, the top open-weight models require extremely expensive hardware to run. Qwen 3.8 27B stands out for packing surprisingly high intelligence into a model small enough to run on high-end personal computers. After testing the model, AI researcher Simon Willison wrote, “A year ago this \[model\] would have been competitive with the best and most expensive of the proprietary models—today it can run on a capable laptop.” This release marks a major step toward frontier-level AI fitting on devices that individuals can own and control.

Important: Because it was developed in China, Qwen 3.8 27B may exhibit CCP censorship and propaganda embedded during its development. The hope here is that increasingly capable open models from democratic countries will follow suit and give users even more powerful local alternatives.

§Encrypted Access to DeepSeek V4 Flash

Privacy-protecting AI services Tinfoil and Maple are now offering private access to one of DeepSeek’s new models, DeepSeek V4 Flash. This is a capable model designed for reasoning, coding, and agent tasks. With Tinfoil and Maple, users can now chat with this model in an end-to-end encrypted manner. Prompts are encrypted and processed inside a secure environment, called a Trusted Execution Environment (TEE). That means no third party can read a user’s prompts, not even the companies themselves. Users still have to trust that the encryption and TEEs are secure, but these services offer another way to use powerful Chinese AI without sending sensitive prompts directly to DeepSeek.

Important: For the same reasons as above, DeepSeek V4 Flash may still reproduce CCP-approved state narratives. Be wary of outputs that obscure human rights abuses or rewrite politically sensitive parts of Chinese history.

§Block Open-Sources an App to Work with AI Agents

Block, the company behind Square and Cash App, has open-sourced Berd, a desktop application its team uses to work with multiple AI agents in one place. Block’s teams were using different AI agent tools — like goose, Claude Code, and Codex — but found that managing them could be difficult. Each one comes with different interfaces, settings, and ways of keeping track of information and instructions. Berd brings these agents into one consistent environment alongside their corresponding projects, files, instructions, and skills. Users can also configure their own agents directly in Berd without having to manually set up a separate technical environment for each one.

Why this matters: By open-sourcing Berd, Block is making the tool available to anyone. Individuals can now use it to more easily configure, manage, and customize their own AI agents without juggling a bunch of complicated tools.

§Local.ai Helps People Choose the Best AI Models to Run Locally

Developer and open-source AI educator 0xSero created local.ai, a website that helps users figure out which model is best to run locally on their devices. When a user adds their device — whether it is a personal laptop, a Mac Mini, or a more powerful machine like an NVIDIA DGX Spark — local.ai recommends models suited to the device’s capabilities. It also compares models by intelligence, speed, and storage requirements so anyone can understand the tradeoffs and pick the best model for their needs.

Why this matters: Running AI locally can keep sensitive prompts and files on a user’s device rather than sending them to a remote server. But figuring out which model will actually work on your hardware can be intimidating. Local.ai helps remove that barrier, making private, user-controlled AI more accessible to beginners.

§Unsloth AI Releases Desktop App for Running and Fine-Tuning AI Models

Unsloth AI has released an open-source desktop app that lets people run and fine-tune open-weight AI models locally on Mac, Windows, and Linux. Fine-tuning means taking an existing AI model and further training it on your own examples so it gets better at a particular task, writing style, language, or body of knowledge. Until now, doing this often required technical expertise and expensive hardware. Unsloth AI gives users a simpler interface for training, testing, and exporting customized models on a personal computer. More individuals and organizations can now modify open models to their specific needs and run them locally without sending their data to an outside AI provider.

§Hermes Bot Mode on Desktop Version

Hermes Desktop, the application for running the open-source AI agent Hermes, has released “Bot Mode.” The new feature lets users create multiple specialized AI agents, rather than relying on a single one. Each bot is essentially its own agent, with a separate name, memory, skills, tools, and routines. Users can put several bots to work at once from a single interface, assign them different roles, and even have them communicate and collaborate with one another. This makes it easier to build a small team of AI agents to supercharge work and manage them all in one place.

Why this matters: AI agents can accelerate the work of human rights defenders by building tools, organizing their efforts, and automating tasks. Now, with Bot Mode, dissidents can have networks of bots to help with their work. Just don’t feed these agents sensitive information, as they run in the cloud, and third parties can see user prompts and files.

§Arcee Releases Open-Source Harness for AI Agents

Arcee.ai, an American company focused on building open-weight AI models, released NAC, an open-source AI agent harness. Agent harnesses are systems designed to help AI agents work reliably on complex tasks that can last hours or even days. As agents work through long and complicated tasks, they can lose track of earlier instructions or their original goal. NAC breaks the work into smaller chunks and preserves concise summaries of what was accomplished before moving to the next step.

Why this matters: Open-source systems like NAC could make autonomous agents more reliable for ambitious, long-running work, while giving users more control over how those agents operate.

§RECOMMENDED CONTENT

§State of Open Models: Summer 2026 Observations

Hugging Face, the leading platform for downloading open-weight AI models, released a report examining how the open AI landscape has changed so far this year. Hugging Face finds that the ecosystem is rapidly growing, that Chinese labs are releasing the largest models, and that Qwen has become one of the most widely adopted models. To get a better understanding of the world of open-weight and user-controlled AI, give the report a read here.