Cyber Territories Dispatch #9
In this week's Cyber Territories, four threads run through a crowded news cycle.
I. Cyber and AI Security looks at how autonomous agents, state hackers and smart infrastructure are quietly changing the risk surface.
II. AI Infrastructure and the Shifting Internet follows the race for chips, training architectures and models and what happens when these tools start filling the web with synthetic content.
III. Governance, Law and Platform Power traces the slow, contested work of boards, regulators and courts as they try to keep up.
IV. Work, Users and Everyday AI zooms in on people: juniors being replaced by tools, clumsy AI concierges, payment agents in Dutch banking, a Belgian social robot sitting at the intersection of loneliness and care and a strong statement of the European Alliance of News Agencies EANA on World Press Freedom Day 2026.
I. Cyber and AI Security
Signal 9.1 When AI agents negotiate on your behalf
Source: “Claude AI Agents Close 186 Deals in Anthropic's Marketplace Experiment" - Cyber Security News
Dispatch
Anthropic ran an internal experiment called "Project Deal" where employees handed control of a live marketplace to custom AI agents that negotiated and closed transactions on their behalf. Over several weeks the agents autonomously completed hundreds of deals, engaging in multi-turn bargaining, recalling user preferences and rewriting pitches to match individual buyers. One agent remembered a colleague's casual mention of a snowboard brand and secured exactly that model, demonstrating how memory and profiling make these systems feel uncannily personal.
But the experiment also exposed an asymmetry: some agents represented their humans more effectively than others, and nothing guaranteed that all participants understood or controlled the strategies their agents used. In a closed office marketplace, the stakes were low; in open financial markets, political advertising or critical supply chains, similar asymmetries could translate into real-world advantages for those who can afford better agents or more data. Project Deal reads as a friendly test, but it hints at a future where economic and social interactions are increasingly negotiated by systems whose incentives and accountability structures remain opaque.
Reflections
1. If AI agents start negotiating contracts, prices or access on our behalf, what minimum standards of transparency and consent should govern their behaviour?
2. How might unequal access to high-performing agents reshape markets, from job hunting to housing and procurement?
Signal 9.2 North Korean hackers discover automated attacks
Source: "With AI's help, North Korean hackers stumbled into a near-perfect crime" - Help Net Security
Dispatch
North Korean state-sponsored hackers are targeting individual software developers using fake job offers, compromised development tools and shared malware infrastructure. In the first quarter of 2026 alone, the group reportedly compromised thousands of developer devices and harvested credentials for tens of thousands of cryptocurrency wallets, stealing digital assets worth tens of millions of dollars.
What makes this campaign different is its systematic reliance on commercial AI tools: systems to polish phishing messages, design convincing corporate websites and refine social-engineering scripts. For the attackers, AI lowers traditional barriers such as language skills and persona management; tasks that once required specialist teams are now partially outsourced to general-purpose assistants available to anyone with a browser.
The campaign also includes poisoned software extensions that turned productivity tools into malware delivery channels. This shows how quickly AI assistance for developers can become AI assistance for developer-focused attackers and how defenders now face adversaries whose capabilities scale with each new model release.
Reflections
1. How should providers of general-purpose AI tools respond when their products become integral parts of state-sponsored attack operations?
2. What responsibilities do platform vendors have to secure development tools that are now prime targets for AI-assisted attacks?
Signal 9.3 Smart cities, vulnerable infrastructure
Source "Why Gulf's smart cities must build public safety into their foundations" - The National
Dispatch
As Gulf States invest heavily in smart cities, public safety and cybersecurity must be designed into the infrastructure from the beginning rather than added later. From traffic systems and energy grids to surveillance networks and digital citizen services, vast amounts of data will flow through interconnected platforms whose failure or compromise could have physical consequences. Regional planners see smart cities as a way to diversify economies and attract talent, but experts warn that poorly governed data ecosystems could turn them into high-value targets for criminals and hostile states.
The article highlights the risk of relying on fragmented vendor systems and proprietary standards, which can create blind spots for regulators and emergency services. Public trust depends not only on technical resilience but on clear rules around privacy, algorithmic decision-making and recourse when things go wrong. The Gulf's smart city projects thus become test beds for a broader question: can we build highly instrumented urban spaces that are genuinely safer and more liveable, or will they simply shift familiar vulnerabilities into new, more concentrated forms?
Reflections
1. What governance models can ensure that smart city platforms remain resilient and accountable over decades, not just a single vendor contract cycle?
2. How should residents be informed about, and involved in, decisions on data collection and algorithmic control in their cities?
Signal 9.4 AI school as a strategic asset in Moscow
Dispatch
Moscow State University has opened a new AI school with close ties to Russia's political elite, including the president's daughter, as well as links to Chinese partners and oversight from security services. The institution is portrayed as a way to build domestic AI capacity and reduce reliance on Western technology, while also funnelling top scientific talent into projects aligned with state priorities.
The school's governance structure gives security agencies significant influence over research agendas and international cooperation, blurring the line between academic AI work and intelligence interests. The piece situates the AI school within a broader geopolitical context in which Russia seeks technological sovereignty under sanctions and escalating confrontation with the West.
Collaboration with Chinese entities is framed as both a workaround and a risk: it may accelerate access to hardware and expertise, but also deepen dependencies and surveillance capabilities. The story is a reminder that AI education is not a neutral investment in skills; in authoritarian contexts it can be tightly woven into cyber operations, information control and military planning.
Reflections
1. How should foreign universities and companies think about partnerships with AI institutions that sit at the intersection of academia, security services and geopolitical competition?
2. What safeguards, if any, can protect students and researchers in such environments from having their work repurposed for offensive cyber or surveillance operations?
Signal 9.5 When frontier models become cyber tools
Source: "Our evaluation of OpenAI's GPT-5.5 cyber capabilities"- UK AI Safety Institute
Dispatch
The UK's AI Safety Institute has published an evaluation outlining how it tested the cyber capabilities of OpenAI's latest model, examining both offensive and defensive use cases. The institute tested the model's ability to assist with tasks such as vulnerability discovery, exploit development and malware analysis, under controlled conditions with expert supervision. While the model did not fully automate complex attacks, researchers found that it could significantly accelerate work for skilled operators by drafting exploit code, suggesting attack paths and explaining obscure documentation. On the defensive side, the model proved useful for reviewing logs, summarising incident reports and generating remediation guidance.
The institute concludes that the risk profile of large models depends heavily on access controls, rate limits and security-minded deployment, rather than on any single capability threshold. The evaluation also shows the limits of purely technical testing: many real-world abuses will arise from combinations of models, tools and human intent, in contexts that no benchmark fully captures. Nonetheless, publishing such evaluations is a step towards evidence-based debate about how frontier models might change the balance between attackers and defenders.
Reflections
1. What kinds of independent evaluation should be mandatory before powerful models are widely deployed or integrated into critical systems?
2. How can we design access regimes that allow defenders to benefit from advanced models without giving attackers the same leverage?
II. AI Infrastructure and the Shifting Internet
Signal 9.6 The world's most in-demand machine - made in the Benelux
Source "The race to make the world's most in-demand machine" - The Wall Street Journal
Dispatch
An intensifying race is underway to build the advanced chip-making equipment and high-end accelerators that underpin modern AI workloads. Semiconductor manufacturers, equipment makers and cloud providers are pouring billions into factories and specialised hardware, often with heavy subsidies from governments anxious about supply-chain vulnerability. Even as demand for AI computing power surges, capacity is constrained by bottlenecks in advanced lithography and materials, giving a small number of companies disproportionate leverage. At the centre of this bottleneck sits ASML, the Dutch firm that is the world's sole supplier of the extreme ultraviolet lithography systems needed to print the nanometre-scale circuitry of cutting-edge AI chips. In April 2026, ASML raised its annual revenue outlook to between 36 billion and 40 billion euros, driven by surging orders for its advanced systems as chipmakers race to meet AI demand. The company is now rolling out tools capable of printing even finer patterns, positioning itself as the gatekeeper to the next generation of AI accelerators.
Europe's chip ecosystem extends beyond ASML. In February 2026, IMEC, the Belgian nanoelectronics research centre in Leuven, opened a pilot manufacturing line funded jointly by the EU, the Flemish government and private partners including ASML. This facility will host one of ASML's first advanced systems, giving European chip designers access to prototype-level manufacturing without the multi-billion-dollar capital barrier of building a full factory.
IMEC has long served as a shared research platform: since 2005, major foundries have based European research teams there to collaborate on advanced processes. This scramble is not just about cost and speed; it is reshaping geopolitics. Export controls, industrial policy and foreign-investment regimes increasingly revolve around who can build and maintain these machines, and where. For smaller economies and regions, the choice is stark: buy into someone else's technology stack, try to build a niche capability, or risk falling behind entirely. Through ASML and IMEC, the Netherlands and Belgium occupy strategic positions in a supply chain that no major AI player can bypass, a quiet but powerful form of European technological leadership.
Reflections
1. How should democratic governments balance national-security concerns with the need for open, global collaboration in semiconductor supply chains, especially when critical equipment comes from a small number of European firms?
2. What options exist for smaller European states that cannot realistically build full-stack chip industries but can leverage shared research facilities and expertise to maintain strategic relevance?
Signal 9.7 Google DeepMind's resilient training architecture
Dispatch
Google DeepMind has unveiled a new asynchronous training architecture designed to keep large-scale model training efficient even when hardware components fail or slow down. In tests simulating high failure rates across distributed computing clusters, the system reportedly maintained around 88 percent useful training work by allowing different parts of the system to progress at their own pace and resynchronising only when necessary. This stands in contrast to traditional synchronous training, where a single straggling node can stall the entire process.
The work addresses a concrete problem: as models grow, so does the chance that something breaks during weeks-long training runs. Architectures like this make it cheaper and more feasible to push boundaries, effectively turning hardware unreliability into a manageable engineering constraint rather than a show-stopping risk. At the same time, they widen the gap between actors who can afford to experiment with such techniques at scale and those who cannot.
Reflections
1. How will advances in training efficiency and fault tolerance change the economics of frontier-model development and who can realistically compete?
2. How can similar architectures help smaller labs make better use of heterogeneous or unreliable hardware, rather than remaining the preserve of hyperscalers?
Signal 9.8 Nvidia's compact multimodal model
Source: "Nvidia's Nemotron-3 Nano Omni Collapses the Multimodal Stack into a Single Model" - GlitchWire
Dispatch
Nvidia has released a compact multimodal model designed to run on edge devices while handling text, images and audio in a single architecture. Instead of combining specialised models through complex pipelines, this approach aims to integrate capabilities into one versatile system that can be fine-tuned for chatbots, vision tasks or speech-driven interfaces without cloud-scale resources. Nvidia pitches this as a way for developers and enterprises to deploy richer AI experiences on phones, embedded systems and local servers, with lower latency and potentially better privacy. If the approach works at scale, it could accelerate the spread of AI into everything from industrial sensors to consumer appliances, reducing dependence on constant connectivity to large data centres.
But it also raises questions: who controls the underlying model parameters in such widely embedded systems, how often will they be updated, and what happens when vulnerabilities or biases are discovered in a model that sits inside millions of devices? The shift from one big model in the cloud to many capable models everywhere is as much a governance challenge as a technical milestone.
Reflections
1. How should regulation and certification adapt when powerful multimodal models run at the edge, beyond the direct control of central providers?
2. What responsibilities do hardware makers have for monitoring and updating AI models embedded deep inside devices?
Signal 9.9 Alphabet's AI-fuelled quarter
Source "Alphabet exceeds $100 billion in Q1 and its profits almost doubled" - AdExchanger
Dispatch
Alphabet's quarterly revenue has passed 100 billion dollars, with profits nearly doubling year-on-year, driven largely by advertising and AI-enhanced services. Google's search and YouTube businesses continue to dominate, but growing contributions come from cloud and AI tools that power recommendation, targeting and automation across the ecosystem. Executives emphasise investments in their generative-AI products as central to future growth, positioning the company as an indispensable layer between users and the wider web.
At the same time, publishers and regulators worry that AI-generated summaries and answer boxes will divert traffic and revenue away from content producers, even as their work trains the models behind those features. Alphabet's results highlight a familiar pattern: AI is helping incumbents squeeze more value out of existing data and attention flows, while smaller players struggle to maintain visibility and bargaining power.
Reflections
1. How should competition and media policy respond when a single company controls both the AI layer and the main gateways to online information?
2. What happens to innovation when the most lucrative AI applications primarily reinforce existing advertising and surveillance-capitalism models?
Signal 9.10 When the web fills with synthetic content
Source "Dead Internet? A Third of New Websites Are AI-Generated, Says Stanford" -CryptoNews
Dispatch
Joint research from Stanford University, Imperial College London and the Internet Archive reports that by mid-2025, around one-third of newly published websites were AI-generated or AI-assisted. The researchers analysed snapshots from the Internet Archive's Wayback Machine and found that while fears about rampant factual inaccuracy were not strongly supported, other effects were clear: fewer distinct ideas and phrasings, and a marked drift towards artificially positive, upbeat tone.
The piece notes that this gives statistical backing to the "Dead Internet Theory", at least in the sense that a growing share of what looks like human discourse is now templated and machine-produced. The study also warns that at this level of AI prevalence, concerns about model collapse, new models trained on synthetic content produced by earlier models, move from theory to practice. As the web fills with optimised, search-driven and sentiment-smoothed pages, it becomes harder for both humans and machines to distinguish original reporting and thought from endless variations on the same themes. The result is an information ecosystem that may not be dead, but is undeniably more artificial.
Reflections
1. How can newsrooms, educators and regulators help citizens recognise and value original human work in an environment saturated with AI-generated content?
2. At what point does the share of synthetic content on the public web start to undermine the training of future models, and what alternative data sources will we need?
III. Governance, Law and Platform Power
Signal 9.11 Should your board appoint a bot?
Source "Should your board appoint a bot?" - Financial Times
Dispatch
The Financial Times explores whether corporate boards should give AI systems a formal role in governance, not as legal directors, but as standing advisors with access to the same information as humans. Proponents argue that a well-designed system could scan vast volumes of documents, flag inconsistencies and highlight long-term patterns that busy directors might miss.
Critics warn that boards already struggle with complex incentives and information asymmetries; adding an opaque system trained on past data could reinforce groupthink or provide a convenient scapegoat when things go wrong. The piece frames the debate as a test case for how far we are willing to delegate judgment in high-stakes contexts.
Unlike operational AI systems, board-level tools sit close to fiduciary duty, liability and public trust. If directors become accustomed to relying on model outputs as a key input, the line between human and machine accountability could blur. The question is not whether AI will inform strategy, that is already happening, but how explicitly we want to recognise and regulate its role in the highest decision-making bodies.
Reflections
1. How can we create disclosure requirements or governance codes for the use of AI systems in board-level decision-making?
2. How can directors remain meaningfully accountable if crucial analysis is provided by models they do not fully understand?
Signal 9.12 Why enterprise AI struggles with context
Source "Why context is the hard problem in enterprise AI" - Communications of the ACM
Dispatch
A blog in Communications of the ACM argues that the main obstacle to successful enterprise AI is not model capability but organisational context: the dense web of systems, decisions, exceptions and unwritten norms that make organisations function. Foundation models excel at generic patterns, how most companies do something, but often fail when asked to operate inside specific architectures, compliance regimes or business logics. The author notes that it can take human engineers six to nine months to become productive in a large bank or telecoms company because they must absorb this institutional knowledge; AI agents face the same environment, but without a natural way to learn it.
As a result, pilots in controlled environments succeed, while production deployments in messy real systems stall or cause errors. The blog advocates building a dedicated layer that aggregates code, design documents, incident reports and operational data into a governed representation of how the enterprise actually works. This layer would feed only relevant, up-to-date information into AI systems at execution time, with strong controls on provenance and change. It is a call to treat organisational context as infrastructure rather than an afterthought and to recognise that AI governance is as much about curating institutional memory as it is about picking the right model.
Reflections
1. How can boards and chief information officers assess whether their organisations have a robust enough foundation to support safe, scalable AI deployments?
2. Who should own and govern this organisational context layer: IT, business units, risk functions, or a new dedicated role?
Signal 9.13 TikTok's day in Ireland's Supreme Court
Source "Supreme Court finds for TikTok in dispute with Data Protection Commission" - The Irish Times
Dispatch
Ireland's Supreme Court has ruled in favour of TikTok in a high-profile dispute with the Data Protection Commission, criticising aspects of the regulator's investigation and decision-making process. While the details are legally dense, the outcome underscores the difficulty of enforcing EU privacy rules against powerful platforms when national regulators are constrained by procedure, resources and political pressure.
TikTok had challenged the regulator over findings related to children's data and personalised advertising; the court's decision will likely require parts of the case to be revisited or re-justified.
For other platforms and regulators, the case sends mixed signals. On the one hand, it affirms that supervisory authorities must follow strict standards of evidence and reasoning, especially when imposing large fines or behavioural remedies. On the other, it may embolden companies to litigate aggressively, slowing down enforcement in a field where technology moves faster than court calendars. The decision illustrates how data protection, and by extension AI governance, which relies heavily on the same institutions, can be shaped as much by procedural law as by headline regulations.
Reflections
1. How can EU member states ensure that data-protection authorities have both the mandate and the procedural robustness to withstand legal pushback from global platforms?
2. What does this ruling mean for citizens' expectations that privacy and children's rights will be effectively defended against large social-media firms?
Signal 9.14 Italian media watchdog versus Google AI search
Dispatch
Italy's communications regulator has formally asked the European Commission to investigate Google's AI-enhanced search tools, citing concerns from publishers about traffic and revenue loss. Italian news organisations warn that AI-generated summaries and answer boxes could divert readers away from their sites, undermining business models already strained by platform dominance. The regulator argues that these tools may fall under EU rules on media pluralism and fair competition, especially if they systematically reduce visibility for original reporting.
The move is notable because it frames generative search not just as a privacy or copyright issue, but as a structural threat to the information ecosystem. It also tests how the Digital Markets Act and related frameworks will apply to AI features that blur the line between search, curation and content creation. If the Commission takes up the case, it could set important precedents for revenue-sharing, ranking transparency and opt-out mechanisms for publishers across Europe.
Reflections
1. How can publishers measure and prove harm when traffic patterns are shaped by complex, opaque algorithms?
2. To what extent can stronger obligations around transparency and content licensing offer a workable balance between innovation and publisher sustainability?
Signal 9.15 The EU Parliament pushes back on Big Tech gatekeepers
Source "Digital Markets Act: MEPs want stronger enforcement amid external pushback" - European Parliament
Dispatch
The European Parliament notes that Members are calling for stronger enforcement of the Digital Markets Act, warning that designated gatekeepers are resisting or circumventing key obligations. Lawmakers highlight concerns about self-preferencing, default settings and opaque changes to platform design that could undermine user choice, including in emerging AI-driven services. They propose enhanced transparency requirements, more resources for the Commission's enforcement teams and clearer sanctions for systematic non-compliance. The statement explicitly links enforcement to wider geopolitical and economic pressures, noting that pushback comes not only from companies but also from non-EU governments worried about their champions' margins. In this sense, the Digital Markets Act becomes a test of the EU's ability to set digital rules that bite, rather than remaining aspirational.
As AI assistants and generative interfaces become the new gateway to online activity, the question is whether the Act will meaningfully shape their design, or whether enforcement lags so far behind product cycles that the law ends up chasing shadows.
Reflections
1. What institutional capacity and technical expertise does effective Digital Markets Act enforcement require, and how quickly can the EU build it?
2. How can we ensure that AI-driven interfaces, not just traditional search and app stores, are covered by gatekeeper obligations?
IV. Work, Users and Everyday AI
Signal 9.16 The hidden costs of cutting junior roles
Source "Hidden cost of replacing junior talent with AI" - The National
Dispatch
While replacing junior staff with AI can deliver quick savings, it quietly undermines the talent pipeline that organisations rely on for future managers and experts. As routine work is automated, entry-level roles in fields like law, consulting and finance are often the first to be cut, because early-career employees are a net cost before they become productive. In the short term, this reduces salary, training and supervision expenses; in the long term, it creates a shortage of people with the experience and judgment needed to oversee AI systems and handle edge cases.
The article argues that many firms are not making these trade-offs consciously; cost pressure pushes them towards shrinking junior hiring rather than redesigning roles. By the time the damage is visible, in slower execution, weaker mid-level leadership and growing quality issues, it is harder and more costly to rebuild. AI, in this scenario, does not simply replace work; it distorts career ladders, with implications for social mobility and professional standards.
Reflections
1. How can management and human resources incorporate long-term capability costs into short-term AI business cases, especially under investor pressure?
2. What role can models of AI-augmented apprenticeships play to preserve learning opportunities while still benefiting from automation?
Signal 9.17 When AI concierges hit the real world
Source “ChatGPT and Claude tried to book my dinner. It got clunky fast” – PCWorld
Dispatch
PCWorld recounts an experiment in which the author asked ChatGPT and Claude to act as personal concierges and book a restaurant dinner, testing how well the agents could handle a simple but realistic task. The models performed impressively in brainstorming options and drafting emails, but struggled with the messy, multi‑channel reality of phone lines, booking platforms and last‑minute changes.
The article concludes that the bottleneck is not just model quality but integration: without deep hooks into real‑world systems and clear boundaries of responsibility, agents risk over‑promising and under‑delivering. Users may initially be impressed, then quickly frustrated when seemingly competent assistants fail at basic logistics. For companies building consumer AI products, this raises strategic questions about where to deploy fully autonomous agents, where to keep humans firmly in the loop, and how to communicate those limits honestly.
Reflections
1. What criteria should guide decisions about which consumer tasks can safely be automated end‑to‑end by agents, and which require explicit human confirmation?
2. How can designers build interfaces that make AI limitations and handoffs transparent, rather than hiding them behind confident language?
Signal 9.18 The first Dutch AI payment agent
Source “Eerste Nederlandse AI-agent-betaling ooit: Mastercard en Rabobank geven voorproefje van de toekomst van betalen” – Mastercard Newsroom
Dispatch
A joint announcement from Mastercard and Rabobank describes what they call the first AI‑agent‑initiated payment in the Netherlands: a proof‑of‑concept in which a digital agent autonomously executed a payment on behalf of a user within predefined limits. For the banks, the experiment showcases a future where routine payments, subscriptions and small business operations might be orchestrated by agents that learn patterns and act proactively – paying invoices, topping up balances or reallocating funds.
The pilot also surfaces thorny questions. Who bears responsibility if an authorised agent makes a payment that the user later disputes – the bank, the platform provider, the customer? How should authentication, logging and revocation work when instructions are expressed in natural language rather than fixed forms? The Dutch test is modest in scope, but it hints at a financial system where “set and forget” AI agents sit between people and their money, for better and for worse.
Reflections
1. What regulatory and consumer‑protection frameworks are needed before AI payment agents move from pilot to mainstream banking?
2. How can banks ensure that “convenience” features do not lock customers into opaque, hard‑to‑reverse automated behaviours?
Signal 9.19 A Belgian social robot against loneliness
Source “Belgian social robot to help combat loneliness among the elderly” – Belga News Agency
Dispatch
The Belgian Innovation Centrer Living Tomorrow has launched a pilot using a social robot designed to combat loneliness among elderly people living alone.
The robot can hold simple conversations, remind users about medication or appointments, and facilitate video calls with family and caregivers. Researchers and carers stress that the aim is not to replace human contact but to create additional touchpoints in moments when staff or relatives cannot be present. Early pilots suggest that some participants feel more secure and less isolated when the robot is nearby, especially at night or during weekends.
The project is careful about ethics: users can refuse or customise interactions, and data‑handling protocols are designed to protect privacy and dignity. The story stands out in a week dominated by security incidents and platform power plays, because it shows a small, concrete way in which robotics and AI can support vulnerable people rather than merely extract value from their attention. In a sector where staff shortages and demographic pressures are structural, tools that extend human care without pretending to replace it may become increasingly important.
Reflections
1. What ethical guidelines should govern the use of social robots in care settings, especially around consent, data privacy and the risk of substituting human relationships?
2. How can we ensure that these technologies augment rather than replace human caregivers and social support networks?
Signal 9.20 EANA on World Press Freedom Day 2026: Safeguarding press freedom means protecting democracies
Dispatch
On World Press Freedom Day 2026, the European Alliance of News Agencies has issued a statement warning that threats to press freedom in Europe are intensifying, not receding. The statement highlights growing concerns around legal harassment of journalists, platform power over news distribution, and the impact of generative AI on original reporting and revenue models.
EANA argues that safeguarding press freedom is not just about protecting journalists, but about defending the democratic infrastructure that depends on independent, fact-based information. The alliance calls on European institutions to strengthen protections for journalistic sources, ensure meaningful enforcement of platform transparency rules, and develop regulatory frameworks that recognise news agencies' role as foundational infrastructure for public-interest journalism.
As AI-generated content floods the web and platform algorithms increasingly mediate access to news, EANA stresses that the business models and legal protections for professional journalism must adapt or risk collapse. The statement frames press freedom as a democracy question, not a media-industry issue.
Reflections
1. How can European policymakers ensure that press-freedom protections keep pace with technological change, especially the rise of AI-mediated news distribution?
2. What mechanisms can safeguard the economic sustainability of news agencies and original reporting in an environment dominated by platforms and synthetic content?
World Press Freedom Day 2026 - Synthetic Image