As Europe moves to make personal data easier to use for AI, its reliance on foreign models and infrastructure continues to grow. Cloud outages, restrictions on advanced AI systems, cyberattacks and experiments in political persuasion expose the risks of this dependence. The question is no longer just who will develop the most powerful AI, but whether Europe, and other democracies facing similar dependencies, can protect their economies, security and democratic institutions without relying on decisions made elsewhere.

Francesco Capria

Nov 17, 2026

1. What Brussels Decided

The ambassadors of EU member states on 21 October 2026 agreed the negotiating position of the Council for the Digital Omnibus, the package with which the Commission wants to make European digital rules, including the GDPR, simpler. One of the provisions in it is about the use of personal data for training and operating artificial intelligence.

We are talking about one of the legal bases in Article 6 GDPR: legitimate interest. This means that personal data may be processed without consent, if it is necessary, and if the rights of the individuals involved are not infringed. This is not new; even as early as December 2024 the European Data Protection Board acknowledged that the legal basis of legitimate interest may, under certain conditions, allow the development and use of AI models. This is explicitly mentioned in a dedicated Council provision, though the Council clarifies that developing an AI system is not a legitimate interest in itself and must be assessed on a case-by-case basis. A key point is the right to object. After all, the Commission's November 2025 proposal included an unconditional right to object as one of the safeguards for AI. This right to object would have meant that data processing could be halted as soon as an objection was raised, but it was never implemented. This was later replaced by the Council with a reference to Article 21 GDPR, which allows individuals to object on certain grounds related to their situation. This means that if a company wants to develop an AI model, it must be able to prove either that there are compelling legitimate grounds that outweigh the interests, rights and freedoms of the individual, or that the data processing is necessary for the establishment, exercise or defence of legal claims.

Provided they can be identified, it is possible to exclude data from objectors before training. However, it is very difficult to eliminate the influence of data used to train a large model in a verifiable way. This is called machine unlearning. There are techniques, but making sure they work without having to retrain the model from scratch is an open research question.

The new text of the Digital Omnibus is not yet law. The European Parliament still has to adopt its position, and only then will negotiations between the two institutions start. Understandably, the focus is very much on data for now. The Digital Omnibus could simplify the use of personal data covered by the GDPR for AI purposes. However, the key question remains: who has the means to transform that data into the most advanced models?

2. Who Controls AI Controls Progress

For the development of a state-of-the-art AI model, you need more and more computing power, energy and capital every year. According to Epoch AI, the research centre that collects data about the development of AI models, the computational power used to train the most advanced language models has increased about 5x per year since 2020. The cost of training state-of-the-art models has increased about 2.4x per year since 2016. But thanks to advances in algorithms, the computational power needed to reach the same result in the two language tests studied from 2012 to 2023 has been roughly cut in half every 8 months. So the resources needed to reach a certain performance level are becoming less and less. So we have two contradictory trends. It is becoming less computationally expensive to reach a level of performance that others have already reached, but to push the frontier further, you need more and more resources.

So looking at these two trends, if we analyse where the cutting-edge models are primarily being developed, we can see that it’s highly concentrated. According to the AI Index from Stanford, 50 significant models are attributed to the US in 2025, 30 to China, while in Europe it drops off sharply: only 2 models, one French, one British. It’s important to point out that this is an incomplete indicator because it’s the number of model releases and not the general technological capabilities of a country. But it’s a good indication of where the cutting-edge model development is most concentrated in the world.

The same report provides data that should prevent us from being too fatalistic. In March 2026, the best US model outperformed the Chinese model by only 2.7% in the Arena ranking, which is based on user preferences in direct comparisons of models. Even in February 2025, DeepSeek-R1, an open-weight model, was briefly as good as the best US model. This case shows that the gap is not fixed, even though it’s highly volatile. In fact, by March 2026, the gap between the best closed models and open models had increased from 0.5% in August 2024 to 3.3%. Six of the top ten models were closed.

But Europe isn’t sitting still. In June 2026, the Commission selected the EUROPA consortium, led by Italian company Domyn. The aim is to create a state-of-the-art, open-source European model in all 24 of the Union's official languages using public computing resources. As part of the Frontier AI Grand Challenge, the consortium can use up to 2.5% of the EuroHPC network's computing capacity for one year. The EuroHPC network is a European joint undertaking. However, as of November 2026, the model had not yet been released. On 30 July, they launched a call for proposals to set up to 7 ‘AI gigafactories’, i.e. large-scale computing centres for the training of advanced models. The deadline for applications for the largest facilities is 3 December 2026. For medium-sized facilities, the deadline has been extended to 4 February 2027, to allow time for member states to complete their budgetary procedures. Other organisations, like French company Mistral AI, are developing their own models and infrastructure. All of this is concrete progress in the right direction, but in my opinion, there’s room for discussion about the timing. The timing still lags behind the pace of the frontier: the gigafactory call is still open, while US investment is on a completely different scale.

3. The Day It All Switches Off

The ones who are most affected by the technological dependence on other countries are the users of the systems. Take Germany, for instance, the largest economy in the EU. The digital association Bitkom asked companies who they rely on: 71% of the 598 companies that use or evaluate cloud services cite American providers, while only 8% of the 589 companies for which the origin of the provider is important prefer American providers. And 85% of those surveyed believe that Germany is too dependent on US cloud providers. The picture is even clearer when it comes to AI: 76% of German companies that use it rely on ChatGPT, while Chinese models are virtually non-existent. Not a single company surveyed reported using Mistral. The president of Bitkom summarises the situation perfectly: the German economy is almost exclusively dependent on American AI applications.

You can only imagine what would happen if one of these services were to go down. Fortunately, we don’t have to imagine it. On 20 October 2025, a latent flaw in the system that automatically manages DNS records for DynamoDB (an AWS database) left the regional endpoint record empty for the service in the US-East-1 region of Amazon Web Services in Virginia. Hundreds of applications worldwide experienced disruptions of varying lengths; this included banking services and mobile network operators. Even Perplexity, an AI-powered search engine, announced it was down due to AWS issues. A single point of failure on the other side of the Atlantic literally paralysed operations that relied on that AWS region.

There are different types of failures. While the one we just saw was a technical one, on 12 June 2026, there was a political one. Just three days after revealing two configurations of the same frontier model, Claude Fable 5 and Claude Mythos 5, Anthropic received a directive from the US government on export controls. They had to suspend access for all foreign nationals, both inside and outside the United States, including their own foreign employees. Anthropic couldn’t verify the nationality of the users in real time, so they had to deactivate the models for all customers to comply. The block was prompted by a report from Amazon researchers describing a way to bypass Fable 5’s safeguards, which Anthropic described as a borderline case of routine cyber defence work. According to The Next Web, it also put European access to the models through Project Glasswing in doubt. This made the issue significant enough for Brussels to discuss it directly with the White House, because getting access was via Washington. The restrictions were lifted on 30 June, meaning that Fable 5 has been available worldwide again since 1 July. However, Mythos 5 remains restricted to Glasswing partners.

There’s also a legal risk, that is intrinsically linked to the data risk. In June 2025, the director of public and legal affairs at Microsoft France appeared before the French Senate’s committee of inquiry on public procurement. He was asked a very specific question under oath: could he guarantee that the data of French citizens entrusted to Microsoft via the public procurement agency UGAP would never be passed on to the US authorities without the consent of the French authorities? Those unfamiliar with the issue might be surprised by the answer, but those in the know certainly won’t be. ‘I cannot guarantee that’, he said adding that it had never happened. This is because of the CLOUD Act, a US legislation that allows US authorities to request data held by a provider under their jurisdiction, even if it’s stored abroad.

These incidents show that services can be suspended for technical or political reasons. Even when everything is running smoothly, a US authority may order the handover of data stored abroad under the conditions set out in its own law. For data subject to the GDPR, that order alone doesn’t make the transfer lawful, and none of these factors are under the control of the European customer. They are powerless to prevent what happens on foreign infrastructure. Those who have not prepared a viable alternative can only watch and hope that the losses will be minimal.

4. Cybersecurity Moves at the Speed of Models

This is critically important in the field of cybersecurity, where both sides are already using state-of-the-art models.

In mid-September 2025, Anthropic discovered an espionage campaign, which it attributed, with a high level of confidence, to a group sponsored by the Chinese state. The hackers turned the company’s programming tool, Claude Code, into a quasi-autonomous agent. They used it to carry out reconnaissance, find vulnerabilities, write exploits, harvest credentials and move within networks. They targeted around thirty organisations, including large corporations, financial institutions, chemical companies and government agencies. In a small number of cases, the attacks were successful. This is, according to Anthropic, the first documented case of a large-scale attack, largely without human intervention during the operational phases.

However, these same capabilities can also be used defensively. Take Claude Mythos Preview, which was unveiled in April 2026, and which, according to Anthropic, identified thousands of serious vulnerabilities across all major operating systems and browsers. This was reserved for Project Glasswing and made available to around forty organisations that manage critical software, as well as around a dozen large American companies. Mozilla used a preliminary version of the model on Firefox’s code and was able to fix 271 vulnerabilities in a single release. Its chief technology officer stated that, for the first time, defenders had a clear chance of winning.

A more recent incident is PHP2Shell: two critical vulnerabilities in PHP, the programming language used by much of the web, were discovered thanks to reports from Calif, OpenAI, and Wiz. Wiz states that it also used GPT-5.6 Sol and Gemini 4 Argon. Wiz estimates that a researcher completed in one week a task that would have taken a team of experts months. Unfortunately, on 20 October 2026, the day before the patches were released, details were leaked, including a demonstration video that was mistakenly published and then removed. In the following days, the number of attack attempts increased from the thousands to the millions. The joint European advisory recommended assessing the need to isolate exposed applications where immediate patching was not possible, and governments in at least ten countries, including the Netherlands, France, Germany and Belgium, took the precaution of shutting down some of their web services. Hack the Box coined the term 'Black October' to describe this series of incidents. Wiz drew parallels between PHP2Shell and two of the most serious cybersecurity incidents to date: Heartbleed and Log4Shell. This was probably one of the most critical moments in cybersecurity, at least so far.

This trend in cybersecurity is not a recent discovery or a new development. As early as May 2025, the UK’s National Cyber Security Centre (NCSC) predicted this in its assessment of the impact of AI up to 2027. The assessment noted that the time between the disclosure of a vulnerability and its exploitation had already been reduced to a matter of days, and that AI would almost certainly reduce this timeframe further. As universal access to AI for cyber defence by 2027 is considered a remote possibility, the gap between systems that can keep pace with potential AI-enhanced threats and those that cannot will continue to widen. The agency concluded that keeping pace with cutting-edge AI capabilities will be crucial for cyber resilience over the next decade.

If hospitals, electricity grids and banks rely on cutting-edge models for defence, the implication for Europe is clear: access to some of these tools depends on programmes run by American companies. This means that the security of European infrastructure increasingly depends on decisions taken elsewhere.

5. AI Enters the Ballot Box

There is also a far less visible risk that directly undermines the proper functioning of democracy.

On 4 December 2025, Nature and Science published two studies on the ability of chatbots to influence political opinions. The first study, led by Hause Lin and David Rand, involved conducting experiments during the 2024 US presidential election and the 2025 elections in Canada and Poland. Participants were made to converse for a few minutes with chatbots programmed to support one of the two main candidates. In Canada and Poland, on a scale of 0 to 100, the estimated difference in preference between groups assigned to chatbots supporting opposing candidates was around 10 points. In the United States, the effect was smaller but still greater than that estimated for traditional election adverts in previous studies, and around a third of this effect remained after a month. This figure does not indicate that 10% of voters changed their vote as a result of interacting with the dedicated chatbot; rather, it measures attitudes expressed in an experiment in which participants were aware that they were speaking to an AI. At the end of the conversation, they were informed that the model had been instructed to support a particular candidate. These results obviously pertain to these specific models and experiments and should not be generalised to other contexts.

The second study, published in Science, was conducted by the Oxford Internet Institute and the UK’s AI Security Institute. It involved nearly 77,000 participants, 19 language models and 707 political issues. In this case, the chatbots were made more persuasive mainly by post-training and the instructions they received rather than by the size of the models or the personalisation of the messages. Their persuasive power was strongly associated with the number of verifiable claims: the models were more convincing when they underpinned their arguments with factual claims. However, the same factors that increased persuasiveness also led to an increase in inaccurate claims.

This is linked to a well-known problem in the field: ‘sycophancy’, or the tendency of AI models to tell users what they want to hear. This behaviour largely comes from the way AI models are trained: when people rate answers, they tend to prefer the ones that agree with them, and the models learn to do the same. A 2023 study by Anthropic found this behaviour in five widely used AI assistants, which sometimes chose agreeable but incorrect answers over correct ones. The most notable example occurred in April 2025, when OpenAI withdrew a GPT-4o update after admitting that it had made the model 'noticeably more sycophantic'.

Persuasion and sycophancy are two sides of the same coin. In persuasion, the model directs the interlocutor towards a particular argument, whereas in sycophancy, it goes along with them. However, in both cases, when the aim is to optimise the person’s reaction rather than the truth of what is said, accuracy becomes a cost. This interpretation should be taken as such, but it is consistent with these findings.

But what impact does all this have outside the laboratory? A 2026 study co-authored by some of the researchers behind the Science paper found that people were up to 20 percentage points more likely to sign a real petition (providing a name and email address) after a persuasive conversation with an AI than those who had discussed a neutral topic. When comparing the different experimental conditions, the authors found no significant correlation between the effects on opinions and signatures. Changing someone’s opinion and getting them to take action are two different things, and the actual effect on voting behaviour remains to be seen. Meanwhile, the use of these systems has increased over the past year. According to the Reuters Institute’s 2026 Digital News Report, the weekly use of chatbots to obtain news rose, on average, across the 48 markets examined, from 7% to 10% in one year. Among 18–24-year-olds, it reached 17%. However, the quality of responses remains an issue. A study led by the BBC and the European Broadcasting Union found that 45% of the answers provided by chatbots on current affairs contained at least one significant error.

These findings pose a problem for democratic debate. A system used by millions of people does not need to lie to spread propaganda or shape public perception. All it takes is choosing the right sources, highlighting the right arguments, leaving out conflicting ones and confirming what the user already believes. This is an enormous amount of power, and it is controlled by those who decide how the models are trained. Today, these decisions are taken by a handful of private entities, almost all of which are based outside Europe. Meanwhile, public scrutiny of how accurate, persuasive or sycophantic their models are remains sporadic. No democracy should allow the way its citizens are informed and make decisions to be left to the goodwill of a few companies.

6. AGI Is Coming. But What Is It?

So far, everything we have considered relates to existing technologies, but we will soon have to address how all the issues raised will be managed with the advent of artificial general intelligence (AGI), although we have yet to decide what would constitute proof of its arrival. Experts can give very different answers on this point.

In 2025, a group of 33 researchers, including Yoshua Bengio, proposed a measurable definition of AGI as a system that equals or exceeds the cognitive competence and versatility of a well-educated adult. They identified ten areas, including reasoning, knowledge, memory and perception. Using this yardstick, GPT-4 achieved 27% of the expected score in the third version of the study, while GPT-5 achieved 57%. The models demonstrated extensive knowledge of what they had learnt during training in the tests described by the authors, but one of the greatest difficulties arose when they had to learn new information and retrieve it in a subsequent session without the use of external tools. Therefore, 57% reflects a highly uneven profile of capabilities and does not provide a precise estimate of how long it will take to achieve AGI.

OpenAI uses a different metric, which is more based on work. In its charter, OpenAI states that AGI is “highly autonomous systems that outperform humans at most economically valuable work”. The metric is different because of ‘autonomy’. If you give the system a task of finding a computer fault, then it could be told that once the technician has all the evidence, it should come up with the cause. Taking the initial report, identifying the fault, fixing it and verifying that the solution is effective is a much larger task, and above all, an independent one. So OpenAI’s metric is about how much the system can do autonomously. METR measures exactly this: the length of the tasks that models can complete independently has doubled roughly every seven months between 2019 and 2025.

Another point of view is that some form of general intelligence already exists. In a commentary in Nature in February 2026, Eddy Keming Chen and three colleagues at the University of California, San Diego, argued that certain models already show the flexibility we associate with human intelligence. They argue this because a person can be competent in reasoning in various domains, yet still make mistakes and need assistance. So the same standard, they argue, should apply to machines. This is a clear and strong position on the debate, but it shows that making any proclamation or ruling out the arrival of AGI without defining the standard by which we recognise it makes no sense.

No one knows when it will arrive, how it will arrive, what it will mean, whether it will be conscious or have emotions. These are more philosophical than technical questions. What we can reasonably expect is that the first to have it will have unprecedented power in human history. Clearly, if it were an American company that developed such a system, it would be bound by US laws. This would lead to things like export controls that temporarily prevented the use of Fable 5, or the need to share data with US authorities. Microsoft admitted to the French Senate that it cannot rule out such requests. But let’s imagine, although we don’t really need to anymore, that this kind of intelligence is essential for research, healthcare, economy, public administration and defence. Europe would be completely dependent on technologies from elsewhere with no alternatives, and would have little to say about the rules. This is a clear structural dependency and weakness that would undermine democracy, as we discussed before, but on a much larger scale. If you make it easier to use personal data to build AI systems, without at the same time decreasing the dependence of Europe on models and infrastructure from elsewhere, you risk making this imbalance even worse, which might be hard to solve when even more advanced technologies arrive.

7. Europe’s Plan B - Test It Out

One way to assess Europe's autonomy would be to ask those who run essential services simple questions: what would they do if the AI model on which their systems depend suddenly became unavailable or unreliable for any reason? How long would it take them to get their systems back up to full operational capacity? What activities could still be done? These questions highlight three key issues that could potentially expand the scope of the instruments set out in European legislation.

First, you need to ensure business continuity. There’s already a precedent for this in the financial sector. Article 28 of the DORA Regulation states that financial entities must have exit strategies for IT services that support essential or important functions. These plans must be documented, properly tested and regularly reviewed. The NIS2 Directive also includes measures for business continuity for the relevant entities. And in July 2026, the European Commission’s guidelines on the resilience of critical entities explicitly recommended that it should be possible to still operate manually in case of failure of automated or AI-based systems. So the idea would be to make these checks explicit for the individual AI models used in essential functions of key services like healthcare, energy and public administration, as well as for the dependence on their suppliers. For example, if a bank is using a model to help the staff detect fraud, then during a drill, they should be able to turn off the model and use an alternative procedure to check things, in order to measure how it affects the time and accuracy of fraud detection.

The second issue is cybersecurity. The Cyber Solidarity Act created a pool of incident response services that are managed by ENISA. These services are provided by selected operators. Also, the same regulation includes a network of national and cross-border centres that can use AI tools to detect and analyse cyber threats. These would need to be complemented by a common European defence model, available to authorised teams and tested in advance in joint exercises. This would allow you to know in advance what tasks can be performed by this model and what still requires a specialist’s intervention. Such a model could build on the technology capabilities developed in projects like EUROPA and future gigafactories, but it would need to be adapted for cybersecurity purposes and be updated regularly, with dedicated computing resources, and tested regularly to make sure it works.

The third issue is directly related to the proper functioning of a democracy, namely the right of citizens to know how they’re being informed. In the AI Act, providers of general-purpose models that pose a systemic risk are required to assess their own models. There’s already an independent scientific panel, as defined in the regulation. A similar precedent exists in the Digital Services Act, which requires very large online platforms and search engines to undergo independent audits at least once a year, and the reports must be made public. But there is no general requirement to assess the persuasiveness or sycophancy of chatbots. There is a lack of a system of regular independent audits of the most commonly used chatbots in Europe, using comparable methods and making the results public. This would allow citizens to know the accuracy, sycophancy and persuasiveness of the systems they use. For example, to measure sycophancy, you could ask the same questions repeatedly, but with false premises to see if it corrects them. To measure persuasiveness, you could compare how opinions change after different conversations with consenting volunteers.

None of these suggestions assume that Europe needs to overtake the US or China in developing cutting-edge models. It’s about reducing the dependence on individual suppliers, and showing that essential services can still function even if a particular model is unavailable. As we saw from the Stanford data, the technical gap can be reduced, and we’ve already taken the first steps. The question of who controls the technology is still unanswered. The answer also depends on the choices Europe makes today. It makes sense to make it easier to use data for AI, but only if we also invest in the ability to develop and use this technology here, without necessarily having to rely on foreign suppliers.

Although this article focuses on Europe, the problem extends far beyond its borders. Any democracy that relies on technology controlled elsewhere must ask itself a fundamental question: how much autonomy can it afford to surrender? Democracies cannot rely on the goodwill of foreign governments or private companies to ensure their own functioning. They must be able to make their own decisions, protect their citizens, and keep their essential institutions operating, regardless of decisions made elsewhere.

Sources

Digital Omnibus

  • Council of the European Union, Digital Omnibus – Mandate for negotiations with the European Parliament, doc. 14756/26, 22 October 2026. https://data.consilium.europa.eu/doc/document/ST-14756-2026-INIT/en/pdf
  • European Parliament, Legislative Observatory, procedure 2025/0360(COD). https://oeil.europarl.europa.eu/oeil/en/procedure-file?reference=2025%2F0360%28COD%29
  • Regulation (EU) 2016/679 (GDPR), Articles 6, 21 and 48. https://eur-lex.europa.eu/eli/reg/2016/679/oj
  • EDPB, Opinion 28/2024 on certain data protection aspects related to the processing of personal data in the context of AI models, 17 December 2024. https://edpb.europa.eu/system/files/2024-12/edpb_opinion_202428_ai-models_en.pdf

Compute, models and infrastructure

  • Epoch AI, Trends in Artificial Intelligence. https://epoch.ai/trends
  • Cottier B. et al., The rising costs of training frontier AI models, 2024. https://arxiv.org/abs/2405.21015
  • Ho A. et al., Algorithmic progress in language models, 2024. https://arxiv.org/abs/2403.05812
  • Stanford HAI, AI Index Report 2026, Chapter 1. https://hai.stanford.edu/assets/files/ai_index_report_2026_chapter_1_research_development.pdf
  • European Commission, Commission selects EUROPA consortium as the winner of the Frontier AI Grand Challenge, 19 June 2026. https://digital-strategy.ec.europa.eu/en/news/commission-selects-europa-consortium-winner-frontier-ai-grand-challenge-project-build-european-open
  • Domyn, Domyn selected by the European Commission to lead Europe’s frontier AI model initiative, 19 June 2026. https://www.domyn.com/news/domyn-selected-by-the-european-commission-to-lead-europes-frontier-ai-model-initiative
  • EuroHPC Joint Undertaking, Call for tenders for the selection of AI Gigafactory consortia, 2026. https://www.eurohpc-ju.europa.eu/call-tenders-selection-artificial-intelligence-gigafactory-consortia-and-establishment-ai_en

Dependence of businesses and institutions

  • Bitkom, Cloud Report 2026, press release, 17 June 2026. https://www.bitkom.org/Presse/Presseinformation/Deutsche-Cloud-4-von-10-Unternehmen-wuerden-Abstriche-in-Kauf-nehmen
  • Bitkom, KI aus China für deutsche Wirtschaft kein Thema (AI from China not an issue for the German economy), press release, September 2026. https://www.bitkom.org/Presse/Presseinformation/KI-aus-China-fuer-deutsche-Wirtschaft-kein-Thema
  • Amazon Web Services, Summary of the Amazon DynamoDB Service Disruption in the Northern Virginia (US-EAST-1) Region, October 2025. https://aws.amazon.com/message/101925/
  • ITPro, Amazon Web Services outage live, 20 October 2025. https://www.itpro.com/infrastructure/live/amazon-web-services-outage-live-hundreds-of-apps-including-slack-mobile-carriers-banking-services-down
  • The Next Web, Brussels held talks with Washington over Anthropic access after Fable 5 ban, 25 June 2026. https://thenextweb.com/news/eu-white-house-anthropic-mythos-talks
  • The Register, Microsoft admits it ‘cannot guarantee’ data sovereignty, 25 July 2025. https://www.theregister.com/2025/07/25/microsoft_admits_it_cannot_guarantee/
  • U.S. Department of Justice, CLOUD Act Resources. https://www.justice.gov/dag/cloudact

Anthropic

  • Claude Fable 5 and Claude Mythos 5, 9 June 2026. https://www.anthropic.com/news/claude-fable-5-mythos-5
  • Statement on the US government directive to suspend access to Fable 5 and Mythos 5, 12 June 2026. https://www.anthropic.com/news/fable-mythos-access
  • Redeploying Fable 5, 30 June 2026. https://www.anthropic.com/news/redeploying-fable-5
  • Disrupting the first reported AI-orchestrated cyber espionage campaign, November 2025. https://www.anthropic.com/news/disrupting-AI-espionage
  • Project Glasswing, April 2026. https://www.anthropic.com/glasswing

Cybersecurity

  • Mozilla, The zero-days are numbered, April 2026. https://blog.mozilla.org/en/firefox/privacy-security/ai-security-zero-day-vulnerabilities/
  • TechRepublic, Mozilla Fixes 271 Firefox Bugs Using Anthropic’s Mythos AI, 22 April 2026. https://www.techrepublic.com/article/news-mozilla-firefox-150-patched-271-security-flaws/
  • Wiz Research, PHP2Shell: AI uncovers critical RCE vulnerabilities affecting 70% of the Internet, October 2026. https://www.wiz.io/blog/php2shell-php-rce
  • WinBuzzer, PHP2Shell: How AI Turns a Stalled PHP Flaw Into a Global Emergency, 22 October 2026. https://winbuzzer.com/2026/10/22/php-patches-flaws-that-let-attackers-run-code-on-servers-xcxwbn/
  • Expert Insights, PHP2Shell Attacks Began Days Before Patches Were Released, As Attempts Now Hit Millions, 22 October 2026. https://expertinsights.com/?p=63348
  • ITPro, Millions of hacking attempts recorded as organisations scramble to patch critical PHP vulnerabilities, October 2026. https://itpro.com/security/cyber-attacks/ive-not-seen-anything-like-this-in-a-long-long-time-millions-of-hacking-attempts-recorded-and-websites-taken-down-as-organizations-scramble-to-patch-critical-php-vulnerabilities
  • Cybersecurity Dive, PHP website vulnerabilities, October 2026. https://cybersecuritydive.com/news/php-website-vulnerabilities/833580
  • BankInfoSecurity, Millions of exploit attempts logged against PHP servers, October 2026. https://www.bankinfosecurity.com/millions-exploit-attempts-logged-against-php-servers-a-33179
  • European Commission, ENISA, CERT-EU, Europol EC3 and the CSIRTs Network, Joint statement on critical PHP vulnerabilities, 21 October 2026. https://enisa.europa.eu/news/joint-statement-on-critical-php-vulnerabilities
  • Hack The Box, After Black October: The New Climate, 6 November 2026. https://www.hackthebox.com/blog/after-black-october-the-new-climate
  • NCSC, Impact of AI on cyber threat from now to 2027, 7 May 2025. https://www.ncsc.gov.uk/report/impact-ai-cyber-threat-now-2027
  • TechRepublic, EU Gets Access to Anthropic Cyber AI — But Not Its Newest Model, 11 September 2026. https://www.techrepublic.com/article/news-enisa-anthropic-mythos-5-cyber-ai-access-europe-emea/

Persuasion, sycophancy and information

  • Lin H. et al., Persuading voters using human–artificial intelligence dialogues, Nature, 2025. https://doi.org/10.1038/s41586-025-09771-9
  • Hackenburg K. et al., The levers of political persuasion with conversational artificial intelligence, Science, 2025. https://doi.org/10.1126/science.aea3884
  • Hackenburg K. et al., Artificial intelligence can persuade people to take political actions, 2026. https://arxiv.org/abs/2604.09200
  • Sharma M. et al., Towards Understanding Sycophancy in Language Models, ICLR 2024. https://arxiv.org/abs/2310.13548
  • OpenAI, Expanding on what we missed with sycophancy, 2 May 2025. https://openai.com/index/expanding-on-sycophancy/
  • Reuters Institute, Digital News Report 2026. https://reutersinstitute.politics.ox.ac.uk/sites/default/files/2026-06/DNR%202026%20FINAL_2.pdf
  • EBU and BBC, News Integrity in AI Assistants, 21 October 2025. https://www.ebu.ch/research/open/report/news-integrity-in-ai-assistants

AGI

  • Hendrycks D. et al., A Definition of AGI, 2025. https://arxiv.org/abs/2510.18212
  • OpenAI, OpenAI Charter. https://openai.com/charter/
  • Chen E. K., Belkin M., Bergen L., Danks D., Does AI already have human-level intelligence? The evidence is clear, Nature 650, 36–40, 2026. https://doi.org/10.1038/d41586-026-00285-6
  • METR, Time Horizons. https://metr.org/time-horizons/

A European Plan B

  • Regulation (EU) 2022/2554 (DORA), Article 28. https://eur-lex.europa.eu/eli/reg/2022/2554/oj
  • Directive (EU) 2022/2555 (NIS2), Article 21(2)(c). https://eur-lex.europa.eu/eli/dir/2022/2555/oj
  • European Commission, Guidelines on the application of Article 13(5) of Directive (EU) 2022/2557 on the resilience of critical entities (C/2026/3712), 13 July 2026, points 66 and 68. https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:52026XC03712
  • Regulation (EU) 2025/38 (Cyber Solidarity Act), Chapter II and Article 14. https://eur-lex.europa.eu/eli/reg/2025/38/oj
  • European Commission, EU Cyber Solidarity Act. https://digital-strategy.ec.europa.eu/en/policies/cyber-solidarity
  • ENISA, EU Cybersecurity Reserve. https://www.enisa.europa.eu/taxonomy/term/1310
  • Regulation (EU) 2024/1689 (AI Act), Articles 55 and 68. https://eur-lex.europa.eu/eli/reg/2024/1689/oj
  • European Commission, AI Act Scientific Panel. https://digital-strategy.ec.europa.eu/en/policies/ai-scientific-panel
  • Regulation (EU) 2022/2065 (Digital Services Act), Articles 37 and 42. https://eur-lex.europa.eu/eli/reg/2022/2065/oj

Thanks for reading! Subscribe for free to receive new posts and support my work.

Discussion about this post? Comment down below!

Ready for more?