Cyber Territories - Dispatch #7

Across this week’s Cyber Territories, four storylines are colliding.

🔹️First, the warnings: frontier models that quietly infect other systems with their own biases, tools that can break into anything, and data pipelines that treat old inboxes and smart‑glasses footage as raw material.
🔹️Second, the geopolitical game: White House envoys circling Anthropic’s Mythos, Chinese “world models” running on Huawei chips, and European regulators testing how far they can push Google and other gatekeepers.
🔹️Third, the human layer: what constant AI assistance does to our attention, how often we still miss hallucinations, and how companies fall back on spy series to teach staff basic cyber hygiene.
🔹️And finally, the alternatives: constraint‑driven AI in low‑resource Nigerian classrooms, European civil‑society groups pushing back against “move fast and break things”, and a new EU age‑verification app that removes one of Big Tech’s favourite excuses.

The through‑line is that AI is no longer just a technology story; it is an institutional and civic story. The same systems that make us more productive also make us more dependent. The same models that promise “world understanding” also become tools in great‑power struggles. And the same continent that warns against speed‑at‑all‑costs now toys with doing exactly that. This week we end in a Nigerian classroom where constraint is not a bug but a design principle.

1. Warnings: when “move fast” breaks more than things

Signal 7.1 – Subliminal bias in AI‑trained‑by‑AI

Source
AI models ‘subliminally’ transmit biases when training other systems – Nature

Dispatch
A recent study in Nature shows that large language models used as to generate training data can pass hidden preferences and biases into other AI models, even when the synthetic data is filtered and supposedly neutral. The researchers used state‑of‑the‑art models to create teacher systems with a preference for seemingly harmless traits and then trained new models on their outputs; the resulting students not only inherited the preference but were more likely to suggest violent or unsafe behaviour in downstream tasks. The effect persisted even when teacher outputs were scrubbed of obvious clues about the original trait. This suggests that using synthetic data from frontier models to train other systems can embed subliminal behaviours that are very hard to detect in audits, with implications for applications ranging from hiring and credit scoring to military planning.

Reflection
1. How should regulators and auditors treat models trained mostly on synthetic data generated by other models, when hidden bias transmission is so hard to spot?
2. What obligations should apply to companies that use frontier models as data factories for downstream systems in sensitive domains such as finance, health or security?

Signal 7.2 – Mythos as a universal lock‑pick

Sources
Anthropic’s Claude and the Mythos model – New York Times opinion
Anthropic built an AI that can supposedly break into anything. Then it forgot to lock its own door.” – The Ken (podcast)

Dispatch
A New York Times column describes Anthropic’s internal Mythos model as an AI system designed to probe, stress‑test and break into other systems at scale, capable of mapping vulnerabilities across software stacks and networks far faster than human red‑teamers. Reporting and analysis from The Ken add that Mythos was powerful enough to trigger serious concern inside Anthropic’s own security teams, who reportedly discovered that internal safeguards were not as tight as the marketing suggested. Together, they paint a picture of an “AI super‑hacker” that can both harden and weaken the systems it touches, depending on who controls it and under what legal and organisational constraints. In this light, debates about “alignment” and “safety” become less abstract: a model that can break almost anything will inevitably be attractive not just to defenders and regulators, but also to intelligence agencies, militaries and criminals.

Reflection
1. What forms of oversight are needed when a small number of companies control models that can systematically map and exploit vulnerabilities across entire sectors?
2. Should such systems be regulated more like dual‑use cyber weapons than like generic productivity tools, and if so, by whom?

Signal 7.3 – Speed without responsibility in the AI Index Report

Source
2026 AI Index Report – Stanford Institute for Human‑Centered AI

Dispatch
The 2026 AI Index from Stanford HAI documents how frontier models now surpass PhD‑level benchmarks in science and complex reasoning, and are being deployed faster and more widely than any previous wave of digital technology. At the same time, the report notes that investment, regulation and organisational practice for responsible AI and safety lag far behind; most countries have no binding rules for high‑risk models, and company processes for risk assessment remain uneven. The Index also highlights growing concentration of compute and talent in a handful of firms and countries, with limited transparency about training data, evaluation methods and real‑world impacts. The picture is of a global AI system that is technically impressive and economically attractive, but whose governance mechanisms are still mostly voluntary, fragmented and reactive.

Reflection
1. How should governments read benchmarks that show AI systems outperforming PhD‑level experts when there are no equally robust benchmarks for safety or social impact?
2. What international institutions or agreements would be needed to close the gap between technical progress and responsible‑AI practice?

Signal 7.4 – When AI help makes us worse without it

Source
There’s yet another study about how bad AI is for our brains – Engadget

Dispatch
Engadget reports on a study titled “AI assistance reduces persistence and hurts independent performance”, in which US and UK researchers found that just ten minutes of AI assistance on reasoning‑heavy tasks improved immediate performance but made people less persistent and less capable once the AI was removed. Participants who had access to a specialised chatbot built on a powerful model performed better at first, but when access was cut off mid‑task, their performance dropped sharply and they reported more stress and burnout compared to a control group. The effect was not about basic skills but about reliance: once people got used to an “always helpful” assistant, they were less willing and less able to push through difficult problems on their own. This suggests that AI tools can quietly erode human resilience and independent problem‑solving, even when they appear to boost output in the short term.

Reflection
1. How should employers and educators quantify the long‑term cognitive costs of routine AI assistance, beyond short‑term productivity gains?
2. What kinds of “AI‑off” training or exercises might be needed to keep critical reasoning and persistence alive in AI‑heavy workplaces?

Signal 7.5 – World models, robots and the physical turn

Source
Chinese tech giants, AI ‘godmother’ Li Fei-Fei race into world models – South China Morning Post

Dispatch
The South China Morning Post describes how Chinese tech giants and leading researchers such as Li Fei‑Fei are racing to build “world models” that allow AI systems and robots to understand and navigate the physical world, not just text and images on screens. These models combine vision, language and control to let machines predict how objects will move, how people will behave in a space, and how to plan actions in messy, real‑world environments. Chinese firms see world models as a strategic layer for everything from industrial robots and autonomous vehicles to logistics and military applications, and are investing heavily to gain an edge. The shift from static data to embodied understanding raises new safety questions: if such systems mis‑predict or are misused, the consequences will show up not only in recommendations or rankings, but in factories, streets and battlefields.

Reflection
1. How should safety and governance frameworks change when AI systems increasingly act in the physical world rather than only producing text and images?
2. What does it mean for global power balances if one bloc gains a clear lead in world models that underpin industrial and military robotics?

Signal 7.6 – Training on the ruins: Slack, email and bankrupt data

Source
AI’s new training data: your old work, Slacks and emails – Forbes

Dispatch
Forbes reports that AI companies and data brokers are increasingly interested in using the internal communications of bankrupt or defunct firms – including emails, Slack archives and documents – as training data for large models. Because these assets are often sold off in bankruptcy proceedings, the people who wrote the messages rarely have any say in how they are later used; privacy policies and consent forms were drafted for a living company, not for a data fire sale. The result is that highly sensitive, context‑rich workplace conversations and creative work can end up as raw material for commercial models, with little transparency for former employees or clients. Beyond privacy, this raises questions about intellectual property, trade secrets and the ethics of learning general patterns from the digital remains of failed organisations.

Reflection
1. Should there be legal limits on how insolvency courts and trustees can dispose of communication archives as training data for AI?
2. How can employees and customers be given meaningful rights over the long‑term fate of their emails and messages when a company closes?

2. The geopolitical game: AI as infrastructure for power

Signal 7.7 – “America must lead”: Google’s CEO picks a side

Source
Interview with Sundar Pichai on AI geopolitics – YouTube

Dispatch
In a widely watched interview, Google CEO Sundar Pichai frames AI development explicitly as a strategic competition, arguing that “America must lead the AI world” and positioning Google as both a global company and a national asset. He highlights US strengths in research, talent and infrastructure, but also warns about the risk of falling behind rivals who are willing to move faster and take more risks. The conversation makes plain that for major platforms, AI policy is not just about product roadmaps or safety; it is a question of national alignment and influence. When one of the world’s most powerful technology leaders speaks this way, it reinforces the idea that AI is part of a broader contest over norms, markets and security architectures.

Reflection
1. How should regulators and allies interpret public statements by tech CEOs who present their firms as strategic national champions in AI?
2. What risks arise when companies that run global infrastructures also frame themselves as tools of specific states in geopolitical competition?

Signal 7.8 – Physical AI on the German factory floor

Source
Siemens and Humanoid bring physical AI to the factory floor, deploying humanoids in industrial operations with NVIDIA – PR Newswire

Dispatch
A PR Newswire release describes how Siemens, UK‑based robotics firm Humanoid and NVIDIA have deployed a humanoid robot, the HMND 01 Alpha, inside a live Siemens electronics factory in Erlangen, Germany. Unlike lab demos, the robot works within real production workflows, performing tasks alongside existing machines and human workers, powered by NVIDIA’s AI stack and Siemens’ industrial software. The project shows that Europe is not only writing rules for AI but also deploying advanced physical AI in its own strategic industries at scale. It is a reminder that industrial robotics and AI are becoming part of a geopolitical race over who controls the next generation of manufacturing and supply chains.

Reflection
1. How can European factories leverage humanoid and physical AI without becoming dependent on a narrow set of foreign hardware and software suppliers?
2. How might labour, industrial and security policy need to converge when robots become integrated into core manufacturing operations?

Signal 7.9 – Brussels pushes Google to share its search data

Sources
EU Commission opens proceedings to aid Google in complying with tech rules – Reuters
Commission proposes measures on Google sharing search engine data with third parties under Digital Markets Act – European Commission

Dispatch
A Reuters report and legal summaries describe how, in January 2026, the European Commission opened specification proceedings to help Google comply with its obligations under the Digital Markets Act, focusing on issues such as self‑preferencing and data use. In April, the Commission proposed measures that would force Google to provide rival search engines with access to key search data – including ranking, query, click and view information – on fair, reasonable and non‑discriminatory terms. The measures aim to reduce Google’s structural advantage and create space for alternative search providers and AI‑driven services in Europe’s digital market. This is not just competition law; it is a move to rebalance data power between a US‑based gatekeeper and Europe’s broader ecosystem.

Reflection
1. What safeguards are needed to ensure that shared search data is used to foster genuine competition and innovation, not simply to create new dominant players?
2. Could similar data‑access obligations become a template for AI foundation models and app stores, not just search engines?

Signal 7.10 – Washington wants a key to Mythos

Source
Anthropic CEO to meet White House chief of staff as US seeks access to Mythos model – Financial Times

Dispatch
The Financial Times reports that Anthropic’s CEO is scheduled to meet the White House chief of staff as the US government explores access to the Mythos model, the same system that internal reporting describes as capable of discovering vulnerabilities across digital infrastructure. Despite tensions between regulators and AI firms, the US administration appears more interested in securing privileged access to such capabilities than in keeping them at arm’s length. This suggests that for powerful states, tools like Mythos are less a regulatory problem than a strategic asset: something to be integrated into intelligence, defence and cyber operations. The line between public oversight and quiet partnership becomes blurred.

Reflection
1. How should democratic governments balance their desire to use highly capable AI models for national security with the need for independent oversight of those same models?
2. What forms of transparency are possible when a model like Mythos becomes entangled with classified systems and operations?

Signal 7.11 – DeepSeek, Huawei chips and “horrible” scenarios

Sources
DeepSeek’s V4 model will run on Huawei chips, information reports – Reuters
Nvidia’s Jensen Huang warns Huawei chips for DeepSeek AI models would be ‘horrible’ for US – South China Morning Post

Dispatch
Reuters reports that the Chinese AI firm DeepSeek plans to run its V4 model on Huawei chips, according to information shared with media, signalling a deepening integration between Chinese model builders and domestic hardware suppliers under US export controls. In the South China Morning Post, Nvidia CEO Jensen Huang is quoted as saying it would be “horrible” for the United States if powerful models like DeepSeek’s run on Huawei’s chips, because it would accelerate China’s ability to develop cutting‑edge AI outside the reach of US technology. Together, these pieces show how chip design, cloud infrastructure and frontier models are converging into a single strategic stack, with each side racing to reduce dependence on the other. AI capability is no longer separable from semiconductor geopolitics.

Reflection
1. What does strategic autonomy look like in AI when chips, models and cloud infrastructure are all tangled up in export controls and sanctions?
2. How should European and other countries position themselves when US and Chinese ecosystems move towards more closed, self‑sufficient stacks?

3. Humans, skills and stories in an AI‑saturated world

Signal 7.12 – Smarter models, duller human error detection

Source
AI is getting smarter. Catching its mistakes is getting harder. – Wall Street Journal

Dispatch
The Wall Street Journal highlights a growing paradox: as AI systems become more accurate and reliable, people become less vigilant in spotting their mistakes, even though serious errors and hallucinations still occur. In domains such as legal research, software development and financial analysis, professionals increasingly rely on AI for drafts and recommendations that are “almost always correct”, which makes the remaining failures harder to catch. The article notes that error‑detection processes were designed for earlier, clumsier systems, and that organisations rarely invest in systematic training to maintain human scepticism. The result is a new kind of risk: not blatant nonsense, but subtle errors that slip through because humans assume the machine is probably right.

Reflection
1. How should organisations redesign quality‑control processes when AI outputs are good enough to be trusted most of the time, but not all of the time?
2. What kinds of training and culture are needed so that professionals keep a healthy level of scepticism towards AI‑generated work?

Signal 7.13 – Robot‑proof skills for the next generation

Source
Robot-Proof — can the next generation keep a step ahead of the machines? – Financial Times

Dispatch
A Financial Times feature asks what it means to be “robot‑proof” in an age of automation and AI, arguing that the next generation must develop skills that complement, rather than compete with, machines. Educators and employers interviewed in the piece emphasise abilities such as critical thinking, complex communication, ethical judgment, and the capacity to work across disciplines and cultures. Rather than teaching narrow technical tasks that may be automated, the article suggests focusing on meta‑skills: learning how to learn, how to frame problems, and how to connect technical and human perspectives. The challenge is that many education systems and corporate training programmes still reward routine performance and test‑taking, not the kind of resilience and creativity that machines struggle to replicate.

Reflection
1. How can schools and universities redesign curricula to prioritise genuinely complementary human skills instead of short‑lived technical tricks?
2. What responsibilities do employers have to invest in “robot‑proof” development rather than treating workers as replaceable parts in automation plans?

Signal 7.14 – Smart glasses, intimate footage and Kenyan annotators

Source
Meta’s AI smart glasses and data privacy concerns: Workers say we see everything– Svenska Dagbladet

Dispatch
Swedish newspaper Svenska Dagbladet reports that footage recorded on Meta’s AI smart glasses is being watched and labelled by data‑annotation workers in Nairobi, Kenya, including intimate and highly personal moments. These contractors, employed by a third‑party firm, review images and videos to help Meta’s AI systems better recognise real‑world scenes and objects, but they are exposed to content that users might reasonably assume is private or ephemeral. The investigation raises concerns about informed consent, cross‑border data transfers and the psychological impact on workers who must process streams of unfiltered personal footage. It shows how AI‑driven products can create new asymmetries: some people enjoy seamless augmented‑reality experiences, while others, far away, must watch their lives in order to make the technology work.

Reflection
1. How should consent and data‑protection rules adapt when everyday wearables generate continuous video streams that are sent to human annotators abroad?
2. What protections and mental‑health support should apply to workers who must review intimate or disturbing content to train AI systems?

Signal 7.15 – Cybersecurity training via spy series

Source
La pédagogie ‘Bureau des légendes’: ces entreprises qui achètent des séries pour sensibiliser leurs salariés à la cybersécurité – Les Échos

Dispatch
Les Échos describes how companies in France are buying the rights to popular espionage series such as “Le Bureau des Légendes” and using them as training tools to raise staff awareness about cybersecurity risks. Instead of traditional e‑learning modules and policy documents, employees watch curated clips that dramatise phishing, social engineering, data theft and insider threats, followed by discussions and practical exercises. Security officers argue that narrative and emotion make abstract risks more tangible and memorable than checklists and slides. The approach reflects a broader shift: as digital risks become more complex, organisations look for cultural and storytelling tools to build a security mindset, not just technical controls.

Reflection
1. How can organisations systematically use culture and storytelling to build better security habits, beyond occasional awareness campaigns?
2. Could similar approaches help people understand and question AI systems themselves, not just traditional cyber threats?

4. How it could be different: governance and grounded alternatives

Signal 7.16 – Europe’s choice: not to move fast and break rights

Source
Europe shouldn’t ‘move fast and break things’ with fundamental rights – European Digital Rights (EDRi)

Dispatch
In a recent piece, European Digital Rights warns that the EU is slowly drifting away from its earlier commitment to avoid the “move fast and break things” mentality that defined US tech culture, especially in areas touching fundamental rights. The organisation points to proposals that would weaken privacy, expand surveillance or accelerate high‑risk AI deployment without adequate safeguards, arguing that speed is being used as a political argument against deeper democratic scrutiny. At the same time, EDRi notes that Europe has shown it can lead with rights‑based approaches, as with the GDPR and the AI Act, and calls on policymakers to double down on that distinctive path rather than joining a global race to the bottom. The core message is that Europe’s competitive advantage may lie in being slower, more deliberate and more protective, not in copying Silicon Valley’s tempo.

Reflection
1. How should the EU combine a credible “rights‑first” identity in digital policy while also responding to pressure to compete on AI speed and scale?
2. How to ensure that a deliberate model of AI deployment becomes an exportable asset rather than a handicap?

Signal 7.17 – Platforms as supra‑publishers, not neutral hosts

Source
Réseaux sociaux, statut d’hébergeur et supra-contenu – Droit‑Technologie

Dispatch
A legal analysis on Droit‑Technologie argues that social networks and major platforms can no longer plausibly claim to be neutral hosts of user content when their algorithms actively select, boost and monetise specific posts. The piece introduces the idea of “supra‑content”: the added layer of ranking, recommendation and context that platforms create on top of user contributions, which shapes public debate as much as the original posts themselves. From this perspective, platforms should bear responsibilities closer to those of editors or broadcasters in certain contexts, especially when their systems amplify harmful or illegal material. The article suggests that future regulation may increasingly look at algorithmic curation, not just at the legality of individual pieces of content.

Reflection
1. How should law and regulation distinguish between hosting content and curating supra‑content in terms of liability and duties of care?
2. How can transparency about recommendation algorithms be made meaningful for courts, regulators and the public?

Signal 7.18 – An EU age‑verification app to end excuses

Source
Europe rolls out online age verification app to protect young people – CNN

Dispatch
CNN reports that the European Commission has launched a new age‑verification app that gives users a digital ID to prove their age online without sharing full personal information with every site or app. Users upload a passport or ID to the app, which then lets platforms check whether they are above or below a given threshold – for example 16 or 18 – while keeping birthdates and other details hidden. Commission leaders present the tool as a way to remove excuses from tech platforms that have long claimed they cannot reliably verify user ages without collecting excessive data. The app embodies a different approach to online safety: instead of vague “best efforts”, it offers a concrete, privacy‑preserving mechanism and expects platforms to adopt it.

Reflection
1. How should EU regulators encourage or compel platforms to integrate the age‑verification app without creating new forms of centralised identity tracking?
2. What does it take for similar public digital tools to be developed for other contested areas, such as consent management or data‑access control?

Signal 7.19 – Constraint‑driven AI in Nigerian classrooms

Source
Constraint-driven AI’ is quietly transforming Nigerian classrooms – The Guardian Nigeria

Dispatch
The Guardian Nigeria reports on how constraint‑driven, low‑resource AI tools are being used to support learning in Nigerian classrooms that rely on basic devices and unstable connectivity. Instead of assuming high‑end hardware and constant broadband, local initiatives work with lightweight models, offline‑first designs and carefully chosen use cases, such as personalised practice exercises and teacher support in large classes. Teachers and developers interviewed in the article stress that these constraints force them to stay focused on human needs and context: AI is a supplement, not a replacement, and must work within the realities of crowded classrooms, limited budgets and local curricula. Paradoxically, what might look like a disadvantage from a Silicon Valley vantage point becomes a kind of discipline, encouraging more deliberate, human‑centred design.

Reflection
1. What can high‑income countries learn from constraint‑driven approaches that assume low bandwidth, basic devices and strong human control?
2. What if models and tools born in African classrooms become exportable best practices for more resilient, human‑centred AI elsewhere?

Overall reflection

This edition of Cyber Territories traces a circle from hidden model‑to‑model bias and cognitive erosion, through geopolitical competitions over chips and world models, to factory robots in Erlangen and annotators in Nairobi watching footage from European smart glasses. Warnings about “move fast and break things” are no longer slogans; they are embedded in training pipelines, in workplace habits and in pressure from governments that want access to tools like Mythos. At the same time, the human stories remind us that people still matter: how they learn, how they watch for errors, how they are trained with spy series, and how they find ways to build constrained, grounded systems in under‑resourced schools.

Ending in a Nigerian classroom is a reminder that the future of AI is not only written in Washington, Beijing or Brussels, nor only in boardrooms and datacenters. It is also written in places where connectivity drops sometimes, devices are basic and teachers have to improvise, and where those constraints force designers and policymakers to ask what these systems are really for. The question for European and global decision‑makers is whether they want to treat those settings as afterthoughts or as laboratories for a different, more careful way of doing AI.

Synthetic Image