![[Report] Keynote: Cybersecurity and Capability Building in the Era of Frontier AI - CODE BLUE 2026 #codeblue_jp #codeblue2026](https://devio2024-media.developers.io/image/upload/f_auto,q_auto/c_limit,w_1200/v1794835744/eyecatch/lwaeb6mt4ljjwzxqdkhi.jpg)
[Report] Keynote: Cybersecurity and Capability Building in the Era of Frontier AI - CODE BLUE 2026 #codeblue_jp #codeblue2026
This page has been translated by machine translation. View original
Hello, this is Usuda.
Everyone, are you doing your security measures? (greeting)
This is a report on the following session held at CODE BLUE 2026.
It was a session in which a speaker from ENISA (the European Union Agency for Cybersecurity) dug deep into how to deal with this AI era and introduced initiatives they are actually carrying out.
Keynote: Cybersecurity and Capacity Building in the Era of Frontier AI
Frontier AI is beginning to fundamentally change the speed and scale of attacks in cybersecurity. The discovery, analysis, and exploitation of vulnerabilities, which used to require advanced expertise and time, are being automated by advanced AI models, and even attacks that combine multiple vulnerabilities are becoming possible.
As a result, conventional vulnerability management and the attack lifecycle are rapidly shrinking. The grace period between the discovery of a vulnerability and its exploitation may shrink from days and hours to minutes, and may even virtually cease to exist. Meanwhile, defenders face a surge in vulnerabilities, AI-generated reports, legacy systems that are difficult to patch quickly, and organizational decision-making processes that still operate at "human speed."
When attackers start operating at "machine speed," how should cybersecurity change?
In this talk, based on ENISA's latest efforts regarding cybersecurity in the Frontier AI era, we will examine the structural changes occurring in vulnerability management, incident response, secure software development, and security operations. Furthermore, we will introduce how defenders can respond through ideas such as rapid risk-based prioritization, automation, AI-powered defense, breach-assumed architecture, and "Cybersecurity as Code."
The basic principles of cybersecurity themselves do not change. However, the "speed" at which we put them into practice may have to change significantly.
Speakers: Christian Van Heurck, Ioannis Agrafiotis
Related Materials
These are ENISA reports that were mentioned in the talk.
- ENISA's view on Cybersecurity in the Frontier AI Era (July 2026)
- Joint Advisory on the Use of Frontier AI Models for Code Review (joint by ENISA and CERT-EU)
Report
- I will report on research at the University of Oxford
- A report on AI-type attacks
- But I am not good at making predictions
- Regarding Agentic AI systems, my colleague was interested in propagation
- I was skeptical about whether an Agentic AI mechanism was really necessary, given that LLMs have hallucinations and jailbreaks
- But in reality, things changed
- Now Agentic AI is widely used
- It is no longer just LLMs; Tools are combined, and they can cooperate toward a common goal
- Threat model
- A threat model regarding the security of Agentic AI
- There are also traditional cybersecurity items
- There are also various activities
- We now also need to be concerned about attacks on the LLM itself
- OWASP TOP 10
- There are LLM threats, prompt injection, and so on
- There are also traditional cybersecurity concerns here
- What we overlooked is that people are using AI to look for better attacks
- Our response is falling behind
- We need to integrate AI into security as well
- The threat situation is bad
- 75% of intrusions can be found in logs, but they are fragmented, so visibility is poor
- From ENISA's report at that time, we learned that various attacks were becoming faster
- And the kill chain is changing too
- Classification of cyber kill chains
- There are traditional ones
- There are also the AI kill chain, reconnaissance, and so on
- ECHO-LEAK ATTACK
- An attack against Microsoft Copilot
- Ultimately exfiltrates data
- It was an instruction to post text to Teams
- There was a mechanism to block this leak
- How do we plot this on the existing kill chain?
- An email is created, sent, and received, and how does Copilot process it
- How does prompt execution occur, and how can data be obtained
- It can no longer be properly mapped to the kill chain
- So we use the AI kill chain
- It includes attack preparation, access to AI systems, and so on
- With the existing one, a lot ends up being done in the delivery part
- Mapping the kill chain in 2D makes it easy to understand
- McKinsey Lili AI Platform Compromise
- SQL injection was carried out
- There were prompts in that database
- Flow of the attack
- Access the API documentation
- Such a flow is hard to map onto the existing one
- It becomes reconnaissance
- AI attacks are hard to map onto the traditional kill chain
- The flow involves finding and checking API endpoints
- The first half tends to be forgotten in AI system controls
- Controlling Agentic AI threats
- Probabilistic controls are not good
- Things like access control should be done deterministically
- Individual controls for Agentic AI are also necessary
- Ensuring visibility into both is important
- What should we do? We have to make architectural decisions
- Can policies be enforced deterministically?
- How do we handle access permissions for access to datasets?
- How do we handle RAG information?
- Recently, the use of Agentic Hubs has been increasing
- How many logs will be generated
- How do we handle the logs
- The challenge is how to collect and interpret logs, since AI tools have to be used
- The way to resolve a crisis is always teamwork
- For AI security, we want deterministic mechanisms
- We are still dealing with problems that have existed for a long time, and our response is delayed
- There are incidents where the latest vulnerabilities take down websites
- It is good that momentum for responding by applying patches is growing
- Every kind of system comes attacking in a suit
- The CVE being used is from 2014
- Systems are compromised within 6 hours
- The era of traditional server security is over
- There is panic, but it will stabilize eventually
- Google/Microsoft are releasing various patches
- Such organizations have budgets
- They can also devote human resources
- But the situation is different on the OSS side
- The security community had pointed this out for a long time, but has not been able to respond sufficiently
- OSS contains high-risk flaws
- On the other hand, many things are accelerating
- Anyone can use AI as a toolbox
- But it is often humans who apply the patches
- And many organizations cannot keep up with patching
- Vulnerability lifecycle
- The number of detected vulnerabilities is increasing
- Detection of zero-days has become slightly more subdued
- Measures for OSS
- A large volume of low-quality AI submissions arrived and they had to be dealt with
- The media focuses on how attackers use AI
- But we must think about using AI in a good way
- At that time such risks did not exist, but frontier models became usable
- We should use them for defense as well
- We must be careful that using vibe coding can introduce various vulnerabilities
- We must also consider the danger of new vulnerabilities being introduced
- We need to change our approach
- About the situation in the EU
- There is the EU AI Action Plan
- What is the EU trying to do?
- Running a hands-on program
- Providing a secure testing program
- Using the AI team to provide guidance, recommendations, and best practices
- Running an OSS support campaign
- There is the EU AI Action Plan
- Running AI-related ground challenges
- Access to frontier models is possible
- Mistral contacted us and joined the framework
- 7-step testing cycle
- Agentic harness
- AI discovery
- Human in the loop
- AI exploit
- Patching
- Cost evaluation
- Response to vulnerabilities
- The world did not end because of the vulnerabilities
- Even if the CVSS score is extremely high, that alone is not enough information
- ENISA's report includes this finding
- 68 vulnerabilities were reported, but whether they are reachable and whether they are exploitable are also needed
- Risk-based actions
- Effective even in the AI era
- Let's do it efficiently
- We will also do canary tests and backup plans
- Investigation of 10 packages
- We focus on those with high importance, but whether false positives are high is a point of concern
- Also pay attention to token costs
- The KEYCLOAK case
- Let's run smart exercises
- How the organization as a whole copes with crises
- Personal information must not be leaked, so it is not enough for only the technical team to work hard
- We created the exercise FIRST LIGHT 2027
- The contents of the document are public
- Social engineering
- We will address this together as well
- We also need to put in place our own security for frontier AI
- ENISA strongly supports OSS
- They are also asking for help
- Let's use AI for its original purpose
Impressions
There is a mountain of things to do, but we need to cover all fronts.
I think the content of the reports is useful, so let's use AI to understand and make use of them (I'll do that from now).