Committee on the Internal Market and Consumer Protection Ordinary meeting
Legend
Transcript markers
- text Sentence being played
- text Search result
- text Intervention cited by the link you opened
Click a sentence to play the video from that moment. Select text to quote it with the speaker, timestamp and link.
More options
IMCO European Parliament
Follow its sittingsDespina Spanou
This only comes into force in December, 2027. And as of then, we will also be able to enforce it fully. But in preparation of the Cyber Resilience Act, we hope that the companies that introduce such models also respect the letter of that act with regard to cybersecurity risks. We need to demystify the myth that this discussion is all about mythos. This discussion is about all those models that bring a change to the way we will do our preparedness.
We share the sense of urgency and are currently assessing together with our European cybersecurity agency, Mister Hans de Vries, the CEO of the agency, is with us today, and he will explain more. We are assessing with ANESA, the cybersecurity authorities of the member states, the impact of Frontier AI on cybersecurity, to take further implementation actions to ensure the protection of our critical infrastructure and supply chains. Our CERTIU, the emergency response team of the European institutions, has already issued an advisory on the topic. And, of course, the European Commission is engaging fully with the company Anthropic on this occasion. It can be other companies when they release their models.
Again, as I mentioned, this is not just about this single brand. It is about all such models that will come, and it is a great opportunity that we have to discuss how we will do that. As part of our technology sovereignty package, the commission will also publish an open-source strategy. This will be very relevant for this topic because it will include measures to prioritize the maintenance and security of open-source software components to address vulnerabilities affecting critical infrastructure. Now let me also share with you some of the key steps ahead of us.
We need to use AI responsibly as a first line of defense against AI and its risks. Models like the one that has prompted all these discussions offer a transformative opportunity to advance the EU's secure-by-design ambition. And to stay ahead, we need rapid adoption of advanced AI tools by businesses and public authorities. And here, we need to be very speedy. The market ready models available today must be deployed without delay to strengthen also our own defenses.
Second, we are working with Eneesa so that Eneesa becomes the global powerhouse for vulnerability management services. We should not forget the power that our European cybersecurity agency already has in identifying vulnerabilities and managing the first database of its kind in the European Union, but also working with the US model in this respect. Lastly, strategic EU investment will be critical to bolstering Europe's AI cybersecurity capabilities, accelerating the adoption across strategic sectors, and reinforcing technological competitiveness. Dear chair, before I close, I would like to take this opportunity to ask for your help as the European Parliament. In fact, as I mentioned, we have a plethora of instruments that already address the issues that we are faced with when it comes to the deployment of these large models and the cybersecurity capabilities that they bring.
What we need is to now see whether we have to adjust the existing legal frameworks. You have in front of you a proposal for the new cybersecurity act that addresses the security of the ICT supply chain. This is an opportunity to see whether, with the help of the European Parliament, we need, if necessary, to adapt the proposal of the European Commission to address any issues that remain. We must also see how the European Solidarity Act with mechanisms such as the European Cyber Reserve or the cyber review mechanism for important incidents can also be used. So these are areas we're also looking at, just as a preview of the work that the European Commission is doing.
And I will pass the floor to my colleague who's the director of the AI office, who will give you also a bit of the technical detail around the deployment of these models.
Lucilla Sioli
Thank you very much, Chair. Good afternoon to all. I'll be very short, simply to remind how the AI act applies in a similar case. I want to highlight the fact and clarify that the AI act only applies at the layer of the model. So it allows us under article 55 to have the information about the fact that certain models will be placed on the market, and also make sure that if these models present a certain risk, these risks are mitigated, and this mitigation is actually checked by third parties.
So that means that we can, through the IAC, check the capabilities of the model. But, of course, the IAC will not have any impact on the capabilities of defense by the users directly. The rules of the AI Act will be enforceable by the AI office under the supervision of the AI office on 08/02/2026. So we are not there yet, but since the rules started to apply, which was August, we have been having discussions and interacted with the different providers of these very large models that we call general purpose AI models with systemic risk because they, like in cyber, we see now they can provide some important risk that need to be mitigated. So we have been having a number of interactions also with Anthropic, which is the company that has developed Mithos.
And Anthropic is a signatory of the code of practice. You may remember that under the AI act, we have developed a code of practice, which was signed by a number of companies to implement exactly these obligations under the AI act, and then Tropic is a signatory of that particular code. Now what we have, in our analysis of the model, we have also seen, of course, that this model presents significant capabilities. And, these leaping capabilities in cyber are actually linked to the fact that also other capabilities of the model have increased significantly. And this, as my colleague was saying earlier, is of concern, of course, because it is not only this model that may have these capabilities, but soon, also, other models will have increased capabilities, and that will put pressure on our capacity to defend.
For example, GPT 5.5 was published by OpenAI not long ago. It is available, but it's basically also vetted only to trusted defenders. So I would simply like to highlight the fact that in reality, these companies are also making available instruments that can be used by defenders. And it would be important, in fact, that this is, you know, a step that is taken. If we can, you know, make everybody aware of this possibility, this would certainly be an important step to be able to defend critical infrastructure and other usages from the capabilities of the most advanced models.
Thank you very much.
Christian Doleschal
Thanks a lot for the interesting information. And now I will give the floor to Hans DeFried from Enesa.
Hans De Vries
Thank you very much. Dear chair, dear honorable members, thank you for the invitation to join this exchange of views. Yes. Advanced AI systems can weaken cybersecurity, but we should also always keep in mind that such systems can also tremendously strengthen cybersecurity. I would like to highlight both sides of AI technology in our meeting today.
And overall, the public debate lately focuses mainly on cybersecurity risks, but we also see the advancement of defense mechanisms. AI is an amazing tool in the hands of people, defenders, and attackers alike, and it's already being used. AI systems can analyze massive amounts of data, identify anomalies, secure security weaknesses, and respond with sophisticated solutions, sometimes solutions that we would, as humans, not really find easily. At the same time, these systems can also introduce new attack surfaces and capabilities for cyber perpetrators of any kind, and in most cases, offer new techniques for same old problems. And this is what the front AI models, not only Mithos, but also 5.5 right now, those models succeeded in accomplishing in practice.
Frontera AI can and should help to optimize systems, solve problems, and even act independently in certain contexts. While such AI technology is incredibly useful, it might also result in a significant cybersecurity concern. Why? Front AI can autonomously locate and help exploit bugs, including long-dormant vulnerabilities in operating systems and web browsers. Therefore, Front AI models are a fundamental shift in the economics of the digital domain.
These models challenge traditional security paradigms and our cybersecurity processes. How? Well, for example, such AI system compressed the vulnerability management life cycle in the attack chain from discovery to exploitation. What we mean with compression is the time it takes from when a vulnerability is discovered in an ICT product until the moment it is exploited by threat actors. And this time frame has gone from years to months, now potentially, it can be minutes.
Accordingly to early assessment by well respected cybersecurity players that have tested Mitos, it showed unique ability to autonomously discover vulnerabilities. We've talked with seven of the 12 that were known, Glasswing users, and they all mentioned that it was a really smart MES system. A zero day vulnerability, which is the vulnerability that is not known before, is a software or hardware security flaw unknown to the vendor, meaning that they have zero days to fix it before attackers exploit it. And these vulnerabilities are highly dangerous because no patch solution currently exists. But it's also a possibility to use it in a positive sense, and I've heard and seen that already.
Large companies use these models and earlier models, which are not so sophisticated, but still show that they use these frontier AI models and earlier models for finding out vulnerabilities in the software before it comes into play. And that is actually security by design and security by default, which is the fact that we really need to do. At the outset, let me highlight that Eneesa has not tested Anthropic methods. We have not given access. ANESA, as well as other national cybersecurity agencies from EU member states, have long anticipated the integration of LLMs and AI tools in our IT environments, including vulnerability management and exploitation.
While industry has utilized AI assisted security tools for several years, front AI models introduced a new level of precision, and it's evolving rapidly. And this is great for technology, but also challenges the cybersecurity community and tests our policy frameworks. However, we feel it is not security's doomsday. The past few years, Anis and our expert have noted on many occasions the steadily narrowing window between vulnerability discovery and exploitation. Once a vulnerability is identified by an agent, the transition to a weaponized exploit is no longer a human-only engineering project, but a machine-speed compute task.
What that makes, let's say, a situation even more challenging is that their ability to change findings would go beyond the discovered vulnerabilities. They can creatively reason about application logic, credentials, configurations, and API access patterns. One company mentioned the machine can remember all the 60,000,000 code lines for a system and use that all at the same time while looking at the different possibilities to not only detect the vulnerability, but also help it solve. And I think that is gonna be the future for us. So, yes, there's gonna be a pain situation for a few years, but in the end, it will definitely help us all.
One thing I want to mention is that ENISSA has a unique ability working with the member states on the vulnerability aspect. And EINISA access a CVE route for European entities, which is a mechanism to have the re vulnerability to be reported and then also help solved. And onboarding and training our first CNA, cyber network authority, numbering authority for today is a major milestone for European cybersecurity. It will strengthen Europe's operational contribution to the global CVE program, and it will improve the reliability, timeliness, and coordination of vulnerability handling across the EU. At a time where those frontier AI models are accelerating, vulnerability discovery and exploitation is growing.
Europe's vulnerability management capacity must keep pace and provide trusted operational support to the whole community. And at the same time, Eneesa has opened an open call to give input for the CRA, Cyber Resilience Act, and the way we want to make sure that cybersecurity designed by secure by default is automated in a process. And those questions are out there right now, and can be looked into. Also, Eneesa is already developing the CRA single reporting platform to be launched this year, which will simplify the reporting obligations for actively exploited vulnerabilities. And, yes, we're gonna have a busy time.
I know that. I'd like to finish with a quote that I can't find that we will continue our efforts to strengthen vulnerability management in the EU together with the member states and also supporting the global vulnerability management process in total. And to do that, we need your help.
Christian Doleschal
Thanks a lot. And I now would give the floor to any members who want to ask a question. Okay. We have to organize it. We start with Pablo.
Pablo Arias Echeverría
Yes. Thank.
Thank you very much, sir. I'm going to speak in Spanish.
Thank.
Pablo Arias Echeverría
Thank you very much to the commission and to their speakers for shedding some light on what's happening. We've read a lot in the press. And for some years, we have been talking about artificial intelligence. Until now, we've only really heard positive things about the fact that AI has come to stay. It's going to disrupt the digital world.
And if we're not careful, we'll miss the boat. Of course, we've heard that it also wasn't free of risk. In the field of AI, that moment has come very quickly, given the computational power that it has. So this is opening the door for models that develop faster and faster. We've heard about the practical benefits of AI, but we also know that it can be used to other ends.
And that's what we have been hearing about now, now that some US companies have developed models such as Mithos. It could be used positively or negatively, but I think what we have to deal with today is that the regulation of AI is fundamental for the development of any or all AI models. We've established 4 levels of risk. And I think that these were very sensible in the definition. The Trump administration has accused us of regulatory excess based on fines, with the intention of deterring anybody who wanted to use technology for aims other than those that we approve of, such as positive outcomes for citizens.
So we set out these limits because of our European values. But in the US, they're seeing now that with the use of Mythos, they have a real cybersecurity problem. It would also seem that they've been trying to use the anthropic mythos model for ends related to defense or war, more specifically. So I'd like to know whether we need to take that into account for the future, because I'm sure that what's been done with methos could be done by anybody else, and it's only a question of time and whether they do it more or less successfully. So the risk is there.
What can we do to mitigate that risk and to try to use technology and artificial intelligence for positive purposes? Lastly, mythos could be used for positive events, but we have to be able to detect those that are intended for illicit or criminal purposes.
Christel Schaldemose
I was surprised that I came on as fast. Well, thank you very much for your information. It was interesting and useful. Several questions come to my mind. But maybe the first one, to the European Commission.
You mentioned the Cyber Resilience Act that will enter into force in a couple of years. Do you think it would be possible to make sure that it could begin earlier, since we clearly need it. I think you said December 2027, so maybe a year before, or something like that. The whole area here is so fast that I think that we need to protect ourselves a little bit better. The second thing is that I mean, one thing is meet MITAS and Tropic and the whole situation around that.
But are we capable? Are we able to kind of defend ourselves? Do we have enough money, enough resources, enough skills to develop our own, well functioning, speedy language models that makes us able to protect ourselves better, I think that could be interesting to know. It could also have been interesting to have Entropic here today, but they are not here. So, instead, I will raise my questions to the commission.
But, this is certainly something we have to be more aware of, and I know that you're doing quite a lot of work. But, I think that we need to speed up, what we're doing in the parliament, in the European Commission, and we need to understand that this is very important for our own safety as well. Thank you.
Christian Doleschal
Thank you very much. And I will give the floor to Pipedia Gork. Gork?
Gheorghe Piperea
Thank you. I'm gonna speak Romanian for a better understanding.
I've listened very carefully to all of you, in particular, the last presentation by the commission. And I'd like to mention one comment that comes back again and again, which is that artificial intelligence is beneficial, and we have to make sure that the benefits it brings are multiplied. I'd like to ask, who is AI beneficial for? Is it going to be beneficial for these interpreters who are going to lose their jobs or drivers? Perhaps I can tell you some more risks that we might face.
The language of AI is like a hallucination. There were decisions made in France against some lawyers who used AI to create the documents, they made up the legal citations. And they referred to non-existent case law in their pleadings. So the question is, who is accountable for the client that loses a case because of this hallucination? Secondly, Canadian authorities have pointed out some very interesting cases.
Some workers, both recent and senior, have been using artificial intelligence, so-called phantom AI. It's not still specifically allowed by their employers, but these workers use AI and put solutions proposed by AI into their working documents, and these are also hallucinations. Thirdly, there's a particularly serious problem with regards to security systems, both civil aviation and military. As AI does suffer from hallucinations, it could invent things in the field as well. Are we sure that it's not going to invent conflicts?
And lastly, the most disturbing thing, there's an American company, Palantir, a few weeks ago, they published a sort of democratic charter for artificial intelligence as that company saw it. This is an illusion because there's not really any real democracy inherent in it. AI can produce some opinions, some of which are right and some of which are wrong. So where is the freedom of understanding in that context? So I think these are some of the most dangerous elements and risks posed by AI. The US and China have to do something, and we have to follow their example.
We just don't have the same amount of money here.
Bart Groothuis
From the E3 Committee.
Thank you very much. And also the former NIS two rapporteur and CSA, thank you very much for the things that you've said, both from the commission side and niece Nisa. I would like to start asking about seeing how to handle the governance. If you speak about handling companies like Entropic or OpenAI with their models, that's one. And they say, okay.
Trusted vendors are in, so you could be comfortable with that. Of course, trusted vendors are in. There are two problems. The first, they're all Americans. No European company there.
No European entity. That's my first result. Second, of course, it is much more serious. It's not just about companies. If you look at Kimi k 2.6 or you look at DeepSeek version 4 Pro, I downloaded this weekend, these are only very capable models.
And if you combine them with Agentex swarming, those are equally capable this summer, equally capable of Entropic and others. How will the AI act on open source handle that? I don't see it happening. Who do we say you cannot do this? This is a problem which is bigger than that.
The second thing on n is two and on the CRA, it doesn't entail new products on our market. It only entails new old products, not the legacy. The majority of our businesses, the majority of our institutions have about 30, 40, 60% of legacy software. None of the governance we have in place is there to do anything about the legacy software. So my personal touch on this would be if I were you, but I'm not you, but if I were you, let's just say, get an advice out to all businesses, get rid of legacy software.
It's killing you. This is where a soul typhoon lives. This is where they live. This is where the hackers live. This is where they will enter your premises and hack you.
This is the cyber bloodbath. Mister DeFries says it will not be there, and I hope he's right. I hope you prove me wrong, but not by words, but by action, actually advising companies to get rid of their legacy software, Prior one++. There's no another thing. No Europeans at the table.
Why not? I can tell you why not. Because this is not about who is competent, whether it's an AI office or it's an ESA or we're European. It's about who has the compute power. The UK is on board because the UK AI office has the compute power and the competence to deal with a powerful model like Mythos.
We don't invest in it. I want to make sure who in Europe is my question. Is it the Schwarz Group? Is it the Lido group? Who has the compute power to actually deal with it if we're on the table?
It's my question to the commission. And I want a concrete answer there. On Glasswing, the project to mitigate the mythos, it's very important. I've spoken to many of the participants who were there the first weekend. Some of the companies told me they found many, many thousands, many thousands of vulnerabilities in their software.
The current system we have is for patching five, fifty, maybe 100 patches a year on a product, not thousands at one time. You say there's no cyber bloodbath. Explain to me how this will be managed. How should businesses in Europe be managed? Which guidance will the European Commission give?
And seeing, okay. We will build a CVE database. What is a CVE database, Mr. DeFries, worth? If we're not talking about hundreds or thousands of CVEs rated by the criticality, but it's about 100,000. What is it’s garbage in, garbage out.
It's so big. The CVE database is losing its relevance. That's my biggest concern. And last but not least, also to talk to mister Pepera, the major thing what you said, mister Pepera, about hallucinations, the best thing that my thought is doing is it doesn't hallucinate. So it looks for ten, fifteen seconds if you use ChatGPT.
This looks like three days in a row. And the best thing is it's called alignment with this initial search task. It's extremely well aligned. It just does that task. It doesn't hallucinate, and it goes for it.
This powerful model can disrupt what we're doing. The way we handle cybersecurity, in my opinion, is not up to date, and I was looking for a more substantive answer. I was looking at how to change laws and the way we cooperate with private businesses. And please advise businesses and CSOs on how to deal with this potential crisis that is coming to us. Thank you very much.
Kim Van Sparrentak
Thank you very much. So if it's true with anthropic claims, methos being able to detect all cybersecurity vulnerabilities, then it could pose a major risk if the AI system falls into the wrong hands. But even if half is true, it opens the door to major cyberattacks at scale. Think of banks being hacked, our transport systems being disrupted, and hospitals just blacking out. And if we understand the experts correctly, Methus is just the first model, and many will follow soon.
And with the US government that doesn't believe in regulating AI, we are apparently in the hands of Anthropic to decide whether they think the EU is a trustworthy partner, the likes of Google or not. We cannot simply rely on the goodwill of companies to share such systems as they please. And then Tropic refusing to come to the parliament today to be accountable to European democratic representatives and answer questions is a shame, and it doesn't give me much hope. And, therefore, we need clear actions, clear rules, and, hopefully, a commission that will intervene where necessary. So commission, dear Anissa, are you currently discussing with Entropic to get at least more transparency about access to methos?
What other measures are you taking to protect us all in Europe? And most importantly, do you think the current AI and cybersecurity rules provide enough tools for this? Because what we mainly hear from this commission is that we have too many rules, and that we should deregulate and simplify. Well, I'm not so sure, especially after hearing you now, that what we have in place right now is actually strong enough to face these new challenges, and make sure that our European citizens and companies are safe. Thank you.
The next speaker, Leila Chabi.
Leila Chaibi
Thank you. Good afternoon, everyone. Thank you very much for those presentations. I'm not going to be very original in what I say. It seems like there's quite a consensus gathering around this committee regarding our concerns about anthropic.
We've seen the offensive capacities of its new I model, and it's raising a lot of concerns. So it's a good thing that we're talking about it today. I'm just sorry that Anthropic has not seen fit to attend. It would have been interesting to question the representatives of Anthropic on their decision to reserve last wing to US based companies and organizations. It would also be interesting to talk about the sovereign capacities of the EU in terms of advanced research, vulnerabilities, and cybersecurity.
And, my new colleagues have this very well. I would like to ask about open-source models. We know that it takes six to twelve months to build open source models based on basic models. That means that even if you have an AI act in six to twelve months, there may be open source models, for example, from China used by hackers who don't care about the AI act. As an aside, I agree with my colleague.
Should we turn the AI act into an omnibus? It's true. We need to strengthen it. But back to open source models, in six to 12 months, we could have a number of those even with a recast AI actor. We would not be able to act, including on public software that could be vulnerable.
I know that my questions are not very original, but I think it's important to repeat them. Does the commission consider the EU legislative framework sufficient, faced with these types of offensive capacity AI models? And what do we think about these open-source models that will come into being? And how are you going to have EU sovereign capacity in advanced research and in vulnerability? Thank you.
Marion Walsmann
Yeah. I'll speak.
German. I've got a question concerning the resources needed by ANESA in order to do justice to the requirements of the legislation, the Cybersecurity Act. And then it's not just a question of industrial AI used by businesses. There are AI tools which are sold and implemented in Europe by European businesses. Perhaps you could comment on that.
Josie Kupsta.
Josie.
Josie. Sorry.
José Cepeda
Yes. I'm going to speak Spanish. Thank you very much. To speak frankly, I have the sensation that the evolution of AI and the technological progress is moving at such a speed that it is going to challenge our own institutional capacity to regulate anything. So what I mean is, as we've been saying, we developed a regulation on AI with a very important hole in it, which is military use of AI.
That's a huge omission in our AI Act. So when we're talking about mythos, we're talking about a structure which was designed not with the defense aspect in mind. And we're talking about how to defend ourselves, how to defend our businesses, how to develop an umbrella for European cybersecurity protection, how our institutions are organized, to be able to create this cybersecurity system. And so when we're talking about AI development, we're talking about security and European defense, not only the use of AI for business purposes or for the technological development of our companies. No, we're talking about developing non-agentic AI, which will be autonomous and not subject to any human control, and which will overcome any barrier, and which, by the way, is developing at a record pace and outpacing our ability to pass legislation.
So when we look at some of the speeches of those who have developed, for example, the North American system with a worldwide military defense company such as Palantir, or when you hear what Alex Scott, the CEO of Palantir, says, well, you realize that we are light years away from this response capability. Defensive AI is not defensive. It is offensive. And we don't have enough hands on deck in Europe to be able to answer this. So my first question, people were here explaining the situation both to ANESA and to people from the commission.
But what we're really missing here is a strategic vision for how to conceptualize AI within this model of security and autonomous defense. Other countries are doing so. I don't think that it's just about creating an ethical safeguard for companies. Anthropic, unlike OpenAI, has been participating in collaboration with Palantir and working on its ethical philosophy. They want to protect human life, but other companies are not doing so.
They are strictly involved in development for military use. So we're talking about North American models. We're always talking about North American companies that have a very low level of ethical oversight. But what about Chinese companies that are under the control of the Chinese communist party? When it comes to deep sea, we don't have any information about it, and its potential is possibly even superior to examples such as Methos from Entropic.
So I don't think that this is a minor issue. This is a serious problem, a structural problem that is highly related to the debates that we're having in the parliament on our strategic autonomy and our digital sovereignty, on our technological developments. But what is true is, well, I would like to ask both Anisa and the European Commission, what possibility do you think there is to improve our security on the world stage? We need to ask this question in the context of the military, security, and defense. Thank you.
Christian Doleschal
Merci. And now I give the floor to Dil Kotnick, and then to Andreas Schwab.
Dirk Gotink
Thank you very much, chair, and thank you to the guests, to the committee, commission, and Eneesa. And also, well, I think the message is quite clear. What it feels like is a bit like one of those moments that I would only read in the history books about a stealth technology or about the discovery of gunpowder or even the discovery of weapon systems that are totally revolutionary or when Stuxnet was developed. And I feel that the words of the commission and also ANIZA, they somehow fall short because they talk about our laws and our institutions. But in fact, the whole point is that the people that develop these systems don't refer to laws or institutions.
They are godlike AI developers. They want godlike consequences. And if you get that, then, I think our laws are simply the longer we talk about it, the more time we lose in trying to find real solutions to them. So I would quite agree with mister Chautas, with also mister Zepeda. What is the commission doing to warn businesses of the consequences of this?
Or at least which three questions should you ask yourself as a business if you have, for example, legacy IT systems in order, and what can you do as fast as possible to limit the damage? Because what is clear is that, of course, mythos and, please, the irony of a Greek god or, let's say, Greek like god name of the system is quite funny, to be honest, developed by Americans. What should they do? Because at the moment, these moments in history have unsettled entire economic systems. They can even unsettle political systems.
this
Dirk Gotink
is not just a question of, oh, we have to update or review our AI act or something. This is, like, proper we have to review how our economic system, our social market economy, our liberal democratic order can actually survive what is coming at us. And the fact that it is that Eneesa doesn't have access to it is totally scary, knowing that, of course, this is not the only model that will have these kind of abilities. So please, commission, are we can we escalate a bit on the escalation ladder towards we are actually in a situation where the urgency is very, very high, and we expect the commission also to send those kind of signals. Thank you very much, chair.
Andreas Schwab
You, chairman, dear colleagues. I'd like to split my questions into three with regard to the Anisran Commission. Now the AI Act is based on market surveillance authorities on that principle, and if you look at Methos, the focus is on vulnerabilities in software. So I'm wondering about surveillance of which systems are too risky, and which are used by private companies. And now general purpose AI model with systemic risk under article 51.
Actions can it take as a commission, SNESA, if a model developer refuses to provide the necessary technical documentation, citing national security interests of a third country, which is of importance. Colleagues have been referring to this. And the second element is on a single market and competitiveness. I mean, Anthropics project, Lastwing, the pilot for MITOS is restricted to a selected group of US-based companies and banks. Does the commission see this as a potential barrier to the competitiveness of European companies?
How do we ensure that European SMEs and financial institutions aren't left defenseless against AI-driven threats because they lack access to the same shielding technology? And, I mean, we will have the AI act trial log today. I think that there is a compromise between the Council and Parliament. And therefore, I think we should stop here discussing simplification or other buzzwords. We have to make the rules for this market applicable and available, and with the sense that citizens and businesses at the same time expect?
Thank you very much.
Christian Doleschal
Thanks a lot. And now I would give the floor to the commission, and we start with Tispina Spano.
Despina Spanou
Thank you. Having worked in cybersecurity for some eight years now, and with many of you, actually, we have developed together the ecosystem. I must tell you that, I am very happy to hear that you want to see more. And, as I mentioned in our introductory remarks, I think the European Commission has now put on the table very relevant proposals where US code legislators have a fundamental role to play. But let me also come to the very specific remarks that have been made, because we mentioned already that there is an existing framework.
So, let's start with that. Our cybersecurity framework is designed to be future-proof, and I would like to also make some clarifications. As the members of the honorable members have rightly highlighted, our existing EU and AI and cybersecurity legal framework provides a comprehensive set of risk-based rules to address possible risks, but we need now to implement them properly and clarify certain aspects. The combined application of the AI act and the cyber resilience act will ensure that all AI systems presenting systemic risks and AI enabled products have strong level of cybersecurity, throughout the cybersecurity protection. This includes handling vulnerabilities throughout the life cycle of such AI models.
Now both NIS 2 and CRA include obligations for critical infrastructure operators and software and hardware manufacturers when it comes to vulnerability handling. We do not start from scratch. We don't need to wait for any Mythos or whatever another brand may be called to discover the vulnerabilities. We have rules in place. Europe has been the first part of the world to have uniform cybersecurity requirements for critical infrastructure.
It is a matter of implementing them and not delaying them, and this is where the European Parliament can play a fundamental role. You asked if we can advance the Cyber Resilience Act. We have been fighting for the last two years that under the simplification discussion, we do not postpone, because this was a request that came from certain parts of the market. So we're very happy to hear that you'd like to see it advanced. The good news is, as you know, we cannot do the miracle of advancing and into force, but there are certain aspects of the CRA that come into force earlier, including the vulnerability reporting, that comes into force, I think, already in September, if I'm not mistaken.
So that's much earlier. So that's also good news. On the legacy issue, from a very knowledgeable member of this house that has been part of what we have created, but just to say that, actually, legacy issues are covered both by the CRA and in any case under the NIS, everything is covered, legacy or not. They need to make sure that they address vulnerabilities no matter whether it is legacy or not, when it comes to critical infrastructure. The CRA requires hardware and software manufacturers to implement cybersecurity requirements on products, taking into account their intended use.
They need to follow a security-by-design approach. The NIS too requires some essential and important entities to adopt a wide range of basic cyber hygiene. And for those of you who understand, we're talking here zero trust principles, software updates, device configuration, network segmentation, identity and access management, and user awareness. All that is covered. And, again, we should not underestimate that this will be, if properly implemented by everyone, properly done.
On the CRA, also to say that it will also be very important that the date enters into force, and we do not see further delays. This is why we have been issuing a lot of guidance on the basis of concerns of the industry, to address those concerns, to clarify and interpret so that we do not have to undo what we did the day it comes into force. And that will contribute to security by design in our market. Now there are also certain very interesting topics that were raised.
Despina Spanou
Instance, military, defense. For the first time in our history, in our defense union strategy, we talk about cyber defense. So, yes, maybe we didn't put the military in the AI act. But when it comes to the use of models for cyber defense, there, the dual use will become very important, and that is part of our narrative in our policies today. Then you asked, do we have enough resources for all that?
Again, you have a fundamental role to play. In the cybersecurity act two that is with you since January, which we have not had the opportunity to discuss yet, maybe we can do something to advance this discussion. We are proposing a framework for the security of the ICT supply chain, and there, we can address a lot of any possible gaps that we have. That discussion is where you, as European Parliament, can tell us as Commission where we may have missed something also linked to AI, and it is a unique opportunity that Europe has. Now let us not think that we are behind. You mentioned the comparison to other jurisdictions.
Yesterday, we read in the news a couple of days ago that in the US, they are considering creating the equivalent of an AI board to have an oversight on models that come in the market. Does this remind you of a jurisdiction that has done that through the AI Act? So I think that here, sometimes we lead by example, and we should not underestimate it. But I do agree, and I think as a commission, we will do our best. And our executive vice president has paid a lot of attention to the letter you have sent about how we can use the existing framework and, where possible, reinforce it, implement it better so that it can address these new issues of vulnerability management.
Now on ANESA, we have made a proposal for a new mandate on ANESA, which is now being discussed in the council, and still has not been discussed in this house. Maybe we can do that soon. And we are spending a lot of time discussing whether ANESA can issue threat alerts, and this and that. I think we need to now up this discussion and discuss how ANESA can help us build the capacity, the cybersecurity capabilities we need to help European agencies with their work. There, we have proposed the increase of resources for NESA and for the European Commission because some of the honorable members asked that question, because indeed, it has been a long time since we equipped the European Commission with the resources it needs and the right experts to address these new challenging issues.
I will pass the floor also to my colleague from the AIA office to complete your questions regarding the AIA Act. Thank you very much. And, again, we truly appreciate the sense of urgency that you put on the cybersecurity issues. We take it fully into consideration in the work that we do today, and we look forward to doing more with you to come up with solutions.
Christian Doleschal
Thanks a lot. And now I'll give the floor to Miss Lucilla Scioli.
Lucilla Sioli
Thank you very much. And, I will only reply very shortly on the IACT, also because we have this evening. And so I just would like to remind that the IACT does take into account the different risk, not only the risk posed by the models, but also those posed by systems that have specific objectives, and these are the risk to fundamental rights, safety, and health. Now when it comes specifically to the models, because METOS is a model, I also want to remind that the AIAC does capture models or systems that are developed for dual use. So only those developed only for military would be excluded by the scope of the AI act.
In terms of enforcement, that, as I said, we will have starting on August 2, we will have the possibility not only to request information, but also we will have the possibility to request access to the model for evaluation purposes. We will have the power to request mitigation measures to models that have a very strong risk, and we will also have the power of issuing fines of up to 3% of global annual turnover. In extreme cases, we can also restrict the availability of the model on the European market. We have been discussing with Anthropic, as we said earlier, and we have also presented to member states. We have described the technical capabilities of the model and the impact it may have.
It is true, and that is the whole reason of the discussion today.
for us. Okay.
Lucilla Sioli
It is true that we, in the European Union, are now depending on the availability of these models. Maybe in certain countries, we don't have similarly capable models. DeepSeq is not capable as GPT or the anthropic models yet. Even the latest model, they presented, but we do have companies like, you know, Mistral, for example, which are quite advanced in these developments. And, we are making available compute capacity exactly to stimulate the further development of this model, and we are paying a lot of attention to the development of Frontier AI.
We had a forum on Frontier AI a week ago. Just one last point before closing is that I want to repeat something I said earlier, that I think it would be very important for defense right now to start using the cyber tools that are made available also by these companies. I think this will help a lot, those who need to defend their critical infrastructure. Thank you very much.
Hans De Vries
Thank you very much, and thank you for all the questions that were asked. Let me first reiterate on the remark bloodbath or doomsday. I do understand, and I agree with Mister Frodeus that there's gonna be a hard time with a lot of organizations that will have a lot more attacks and problems with software and hardware than they currently have. So whether we call that a black box or at least it needs a lot of attention. Currently, it is finalizing consulting with the national authorities, the European Commission, as well as organizations producing and accessing models, a paper to understand and to communicate today's operational pressures, tomorrow's strategic pivots, and the day after tomorrow, existential uncertainties and questions.
That document is almost finished. We're gonna talk about that tomorrow in our executive board before it comes out. That's our process. And this collective input will surely help shape Europe's unified response and understanding of what's happening. And in assessing the impact on vulnerability and patch management, but also on security fundamentals, Eneesa is consulting with industry and academia, including manufacturers and SMEs, regarding some of the immediate and midterm cybersecurity challenges and recommendations, and I will just mention a few.
Paper-only security is over. The cybersecurity community knows what needs to be done, and it needs to ensure that organizations move from strategy vision to execution. Zero trust requires commitment and investment. It can be challenging in legacy environments. I agree with that.
A transition needs to be smooth. For example, the EU cyber reserve can offer support to member states in case there is a sharp increase in incidents. So we have a fallback mechanism in place. But surely, cybersecurity as code, machine-speed threats need to be addressed with machine-speed defenses. Vulnerability management as code, instant response as code, cybersecurity by design as code, security architecture as code.
And so we see that that's definitely going to be the possibility. On the second question concerning resources, I'd like to state that the cybersecurity act proposal, which is currently being discussed, seeks to equip Eneesa to be able to better support EU member states and EU to tackle the challenges ahead of us. Perhaps we also need to apply the lessons learned from the Mitros case to the legislation currently undergoing a review or going to be reviewed here soon. We have to improve our common EU cybersecurity capabilities and resilience, and Einisa is committed to help to this end. And that would be my answer.
Christian Doleschal
Thanks a lot, and thanks for all for this interesting discussion. We have different views on this topic, and I also thank the experts from the commission. And now I would close this point and also close the session for today. Thank you very much.
Perfect.
Ask anything about this video — I analyse its transcript to answer.