Welcome to the 22nd edition of HRF’s AI for Individual Rights newsletter.
This week, we examine an investigation into the growing influence of the UAE’s national security adviser, Sheikh Tahnoon bin Zayed Al Nahyan, over the country’s AI future. He has a history of overseeing the country’s surveillance operations, and today presides over a $1.5 trillion business empire. Among his holdings is G42, one of the country’s leading AI companies, raising concerns that its capabilities could strengthen the UAE’s repression of civil society.
Meanwhile, Anthropic launched OSS Scanner, a free AI-powered service that identifies security vulnerabilities in open-source software. OSS Scanner could help Freedom Tech developers audit the code and identify vulnerabilities in the technologies that dissidents and civil society depend on. At the same time, a US-based AI lab, Nous Research, has secured $90 million to expand Hermes, its open-source agent — an investment that could bring more capable open-source agents to human rights organizations and defenders.
We close with an essay from Nicaraguan opposition leader and president of the World Liberty Congress, Félix Maradiaga, who draws on HRF Chief Strategy Officer Alex Gladstein’s talk at Imagine IF to explore how AI closes the vast resource gap between dissidents and dictators. Yet he also warns that these benefits could be lost if powerful AI tools become concentrated in the hands of a few governments or corporations.
The Latest in AI for Repression
UAE National Security Adviser Is Working to Build a Global AI Empire
A recent Guardian investigation reveals how Sheikh Tahnoon bin Zayed Al Nahyan, the UAE’s national security adviser and brother to UAE’s president, is gaining influence over the trajectory of global AI. Tahnoon is a senior figure in a regime with a documented history of surveillance, persecution of dissidents, and other human rights abuses. He oversees an estimated $1.5 trillion financial empire and chairs G42, a UAE-based company building data centers and AI infrastructure. G42 also markets security tools it says offer “suspect profiling,” “relationship mapping,” and “behavioural analysis.” Of equal concern, The Guardian traces G42’s origins to DarkMatter, the UAE surveillance firm that, under Tahnoon’s oversight, ran the Project Raven operation that, according to Reuters, hacked the phones of dissidents and journalists. Former DarkMatter executives, including G42 CEO Peng Xiao, now hold senior positions at G42. Combining the UAE’s surveillance apparatus with G42’s AI capabilities risks giving this authoritarian regime even more powerful tools to expand surveillance and suppress dissent.
In context: Tahnoon has also built close ties to Silicon Valley and Washington: OpenAI’s Sam Altman calls him “a dear personal friend,” Microsoft has invested $1.5 billion in G42, and Donald Trump hosted him for a White House dinner. The Trump administration has also moved to allow the export of hundreds of thousands of advanced AI chips to the UAE, many of them destined for G42. These relationships could help legitimize Tahnoon as a global AI leader despite his connections to the UAE’s repression.
Nigeria Expands AI and Biometric Monitoring Ahead of Its 2027 Elections
Nigeria’s Independent National Electoral Commission (INEC) has created a dedicated AI unit to spot irregularities and combat disinformation ahead of the 2027 general elections. The unit hopes to detect deepfakes and cloned audio of election officials and improve voter verification technology. Details about these plans remain unclear, but INEC separately acquired 57,000 new voter verification devices that use facial recognition and fingerprint scans to confirm voters’ identities. While INEC claims the devices are to prevent election fraud, they also introduce risks to voter privacy and could give officials a way to surveil voters.
Why this matters: Nigeria’s hybrid authoritarian regime has a history of systematic voting irregularities and has been accused by opposition groups of placing ruling-party allies in key INEC positions. Expanding the commission’s AI powers and access to voters’ data could create new opportunities to protect regime narratives, surveil opponents, and tighten political control.
Russia Plans to Double Censorship Budget
According to Russia’s draft federal budget, the regime plans to nearly double spending on internet censorship in 2027, an increase from $115 million to $235 million, with roughly $700 million planned over the next three years. Its censorship apparatus allows officials to block websites, restrict virtual private networks (VPNs), and control which tools Russians can access online — including those crucial to dissident movements. Roskomnadzor has been tasked with blocking VPN traffic with 92% efficiency by 2030 and is reportedly considering, according to Ukrainian intelligence, the use of AI to detect encrypted traffic and automatically block VPN connections. Ukraine’s Foreign Intelligence Service says a separate $1.74 billion is planned for state-controlled media and propaganda in 2027. The combined effort could further isolate Russians from independent news while tightening the regime’s control over what they can read, share, and access online.
India Approves AI Traffic Cameras Across Delhi
India’s cabinet has approved a ₹1,789.52 crore (about $200 million) AI-powered traffic system that will expand automated enforcement to 1,029 sites across 42 corridors in Delhi. This plan would grow Delhi’s network of AI-enabled cameras from about 200 to roughly 2,500. Officials said the cameras will automatically detect red-light jumping, speeding, wrong-way driving, illegal parking, and riding or driving without a helmet or seatbelt, and help identify stolen vehicles and repeat offenders. Privacy advocates have already flagged the plan as “surveillance infrastructure,” noting that a network this size, once built, can be repurposed for far more than traffic.
In context: During recent protests in New Delhi, police deployed AI-powered surveillance tools to scan crowds while also repeatedly restricting the internet, likely to block protesters’ communication. These new AI-powered cameras in Delhi could add another layer to the repression and make it harder for dissidents to organize without being identified or punished.
Chinese Hackers Target US AI Policy Experts
Cybersecurity firm Proofpoint found that a China-aligned hacking group, TA419, impersonated prominent AI policy experts to steal information from researchers at US think tanks, universities, and law firms. The attackers posed as a former White House official, a prominent economist, and a senior Anthropic employee and invited researchers to discuss AI policy. They then sent links to fake Microsoft login pages capable of stealing access to their Microsoft accounts. It is unclear how many, if any, individuals were successfully compromised. Proofpoint assesses the campaign likely supports wider Chinese intelligence efforts to gather information about US AI regulation and export controls.
In context: China is an increasingly dominant force in the AI world, and its authoritarian regime likely wants to continue spreading that influence. Espionage campaigns like these could give state-aligned actors insight that could help advance their AI ambitions.
Recommended Content
Training on the Party Line: Chinese Political Influence on LLMs in China and the World
A new study from Aleph Alpha, a German AI company, analyzed how closely Chinese open-weight models align with the Chinese Communist Party (CCP) stance. Researchers tested six Chinese models with 967 politically sensitive prompts and found that only 17% to 41% of responses were balanced. The rest echoed CCP positions and propaganda or refused to answer. This influence can spread to models built outside China; NVIDIA’s Nemotron Cascade 2 repeated CCP framing in about 17% of researchers’ prompts. The authors caution that as Chinese models generate an increasing amount of online content, their political bias could influence the data that is used to train future AI systems. Read the study here.
The Latest in AI for Freedom
Anthropic to Scan Open-Source Software with Frontier Models
Anthropic, the developer of the Claude AI models, has introduced OSS Scanner, a free service that uses Anthropic’s top AI models to identify security vulnerabilities in open-source software.
Maintainers apply to have their projects scanned, with Anthropic prioritizing those it considers critical to digital infrastructure and user security. Selected projects receive free scans that identify weaknesses in their code and, where possible, suggest fixes. The service could help harden open-source software, including freedom technologies that dissidents and civil society depend on, all without the developers incurring costly AI bills to scan the code themselves.
Important: OSS Scanner’s reports are fully model-generated, without human review, so some findings may be wrong. The scans also run on Anthropic’s systems, so maintainers must trust the company to keep unpatched vulnerabilities secure.
Nous Research Raises $90 Million to Expand Hermes Agent
Nous Research, the AI lab behind the open-source agent Hermes, has raised $90 million “to create AI that serves the people.” Hermes Agent is an open-source AI assistant that has become one of the most widely used, with its code cloned more than 24 million times, according to Nous. These new funds will help develop Hermes for Businesses, a version of the agent built for companies. If deployed correctly, firms could run Hermes on their own servers, keep their data in house, and pick which AI models power it based on cost and task. Although the agent is designed for businesses, civil society organizations or even individuals could adapt this open-source tool to build AI assistants optimized for their work.
In context: Investors include NVIDIA, Microsoft’s venture capital fund, Samsung, Robot Ventures, Union Square Ventures, Y Combinator, and Menlo Ventures. A project that began independently now has major corporate investors. This means commercial interests could influence its development, potentially pushing future features behind paywalls or limiting how open the project remains. That said, Hermes’ current code is open source, meaning that even if Nous changes course, anyone can continue using and modifying its existing code without the company’s permission.
AWS Releases Open-Source Tool to Contain AI Agents
Amazon Web Services (AWS), the cloud computing arm of Amazon, introduced Strands Box, an open-source tool currently in developer preview that runs AI agents in an isolated, or “sandboxed,” environment. It is designed to make every action an agent takes — whether opening a file, running a program, or connecting to a website — comply with rules the user sets in advance. Those rules can be very specific, such as allowing an agent to read a folder without editing it or setting a daily cap on its spending with a paid service. Strands Box works with any AI model and most agent software. But AWS cautioned that potential bugs in Strands Box could allow AI agents to circumvent users’ rules. Sandboxing adds a layer of protection, but it is not a guarantee of security.
Why this matters: Local agents like Hermes can open files, browse the web, and run programs for users, making them powerful but risky on devices with sensitive data. If misused or misconfigured, they could expose contacts, sources, or financial records. Tools such as Strands Box could help mitigate these risks.
Underdog Adds Local Video Generation
Underdog, a local AI assistant developed by US-based lab Conway, has introduced a new video-generation feature called Underdog Videos. Users can generate clips from a text description, animate a still photo, and add sound or music. Conway says everything happens directly on users’ devices, without sending any data to a third party. This works because the AI models powering the feature run locally, with options including LTX-2.5 from Israel’s Lightricks, Wan 2.2 from China’s Alibaba, and Kandinsky 6 Lite from Russia’s Sber. The feature is still in beta, but human rights defenders may one day use it to create social clips without uploading sensitive material to external AI servers.
In context: Launched earlier this month, Underdog is a local AI assistant that handles chat, email, calendars, meeting notes, and now video generation. Although it uses open-weight AI models, the broader software is closed source, meaning its privacy and security claims cannot be verified independently. Users handling sensitive information should remain cautious, including when using Underdog Videos.
Ai2 Releases Open Tool for Training Powerful AI Models
Ai2, an American AI nonprofit focused on building open AI models, has released Olmo-core 3, a new open-source system for training large AI models more efficiently. It is designed to build mixture-of-experts models, a type of model that activates only the parts it needs for each task rather than the entire model every time. That can cut computing costs and make powerful AI more efficient. Because Olmo-core 3 is open source, researchers and smaller organizations can use it to build better open models that could eventually run locally and privately on people’s own hardware.
Recommended Content
AI Superpowers for the Defenders of Freedom by Félix Maradiaga
Félix Maradiaga, Nicaraguan opposition leader and president of the World Liberty Congress, has published an essay exploring how AI brings unprecedented capabilities to human rights defenders. Maradiaga reflects on HRF Chief Strategy Officer Alex Gladstein’s talk at the Imagine IF summit to show that the problem for activists is rarely a lack of courage, but a lack of resources. AI can help close this gap, but Maradiaga warns that using AI safety fears to restrict access could concentrate the most powerful tools in the hands of governments and corporations, leaving activists dependent on systems they cannot control. Read the essay here to learn how the World Liberty Congress plans to put AI into the hands of those who need it most.
