The EU’s cybersecurity agency used an OpenAI model to find four flaws in code for an EU project, Politico reported on 10 September. ENISA spokesperson Laura Heuvinck confirmed the findings to Politico’s Sam Clark. One flaw carries a high-risk rating, and an attacker could use it to hijack accounts. The flaws have since been fixed, Politico reported.
ENISA did the work with CERT-EU, the team that defends the EU institutions’ own systems. According to Heuvinck, the two ran a “security analysis” of the code with an advanced OpenAI model. Politico said nobody had reported this use of the new tools before. TNW has not independently verified the report.
Politico linked the high-risk flaw to CVE-2026-73431. The OpenCVE listing for that record places it in Vulnerability-Lookup, open-source software for tracking vulnerabilities. It affects versions up to 5.5.1.
The software did not track whether someone had already used an activation or recovery token. An attacker with one such link could replay it while it was valid, reset the password, and take over the account again and again. The record scores the flaw 8.8 out of 10 and dates it 12 August.
Europe waited months for access
The EU gained access to the most capable US models in July, after months of requests, Politico reported. The pressure started in April. That month, Anthropic released its Mythos model to a small group of US organisations only.
ENISA later gained access to Mythos. On 10 September, the European Commission said the agency is now testing Mythos 5 and GPT-6 Astra. EU authorities still lack the newest version of Mythos, according to Politico.
On the OpenAI side, ENISA belongs to the company’s trusted access programme for cyber models. The French outlet IT Social reported in June that OpenAI had set up these partnerships with France, Germany, and ENISA, among others. OpenAI runs the access through Daybreak, its cyber defence programme.
Tom Duff Gordon is OpenAI’s head of policy for Europe. In a statement quoted by Politico, he spoke of a “narrowing window” for AI to find weaknesses before attackers do. “That’s why we work with partners such as ENISA,” he said.
Poland’s CERT used OpenAI’s models too
ENISA is not the only European team doing this. CERT Polska, Poland’s national response team, published six flaws in MikroTik’s RouterOS on 5 September. Attackers are chaining two of them, which the team calls MikroTrick. The pair gives full control of routers that expose SSH to the internet. CERT Polska has seen these attacks since at least 2 September.
The team found the flaws with OpenAI’s GPT-5.5-cyber and GPT-5.6-sol models, it wrote. It also set a limit on that claim. The models sped up the analysis, but the team still had to test every hypothesis on real systems. Its own staff judged the impact of each flaw.
A second AI review, on the new CRA platform
A separate AI code review covered ENISA’s own infrastructure. AISLE builds AI tools for vulnerability management and has offices in Prague and San Francisco. On 14 September, it said it had reviewed the code of the Single Reporting Platform for the Cyber Resilience Act.
Hans de Vries, ENISA’s chief cybersecurity and operations officer, is quoted in AISLE’s announcement. He said ENISA had also carried out user and security testing with stakeholders, including national CSIRTs. “I thank AISLE for their important AI-based secure code review performed,” he added. AISLE said it will keep reviewing the platform for the next 12 months.
The platform went live on 11 September. Manufacturers of products with digital elements sold in the EU must now use it to report exploited flaws and severe incidents.
The Register reported that an early warning is due within 24 hours. A fuller notice follows within 72 hours, and a final report within 14 days of a fix. The 24-hour deadline applies wherever a manufacturer is based.
The act counts these reporting duties as core obligations. Breaching them can bring its top fines of €15m or 2.5% of annual turnover, whichever is higher. Most of the act’s other rules apply from 11 December 2027.
What ENISA is building next
ENISA and CERT-EU have used advanced models since July to “actively scan” EU institutions, Heuvinck told Politico. The Commission published an action plan on AI and cybersecurity in July. Under it, ENISA and the Joint Research Centre, the EU’s science arm, are building a “secure testing platform” for the most advanced models.
ENISA is consulting the centre on the design. At the end of September, it will hold a workshop on cyber and AI with the Interinstitutional Cybersecurity Board, which governs CERT-EU.
The agency set out the stakes in a July paper on frontier AI. Attackers can now weaponise a flaw within 15 minutes of its disclosure, ENISA wrote. The paper also said EU organisations need access to AI models and to develop their own. Politico wrote that the agency will treat the four flaws as proof that its long push for early access was justified.
Get the TNW newsletter
Get the most important tech news in your inbox each week.
Facts Only
* ENISA and CERT-EU used an OpenAI model to identify four flaws in EU project code.
* One flaw (CVE-2026-73431) was rated high-risk and allowed for account hijacking.
* CVE-2026-73431 affected Vulnerability-Lookup versions up to 5.5.1 and was dated 12 August.
* The EU gained access to advanced US AI models in July.
* ENISA is testing Mythos 5 and GPT-6 Astra.
* OpenAI provides access via the Daybreak cyber defence programme.
* CERT Polska used GPT-5.5-cyber and GPT-5.6-sol to find six flaws in MikroTik’s RouterOS.
* AISLE performed an AI-based code review of the Single Reporting Platform for the Cyber Resilience Act.
* The Single Reporting Platform went live on 11 September.
* Manufacturers must report exploited flaws within 24 hours under the Cyber Resilience Act.
* Non-compliance with reporting duties can result in fines of €15m or 2.5% of annual turnover.
Executive Summary
European cybersecurity agencies are increasingly integrating advanced US-developed AI models into their vulnerability research and infrastructure defense. ENISA and CERT-EU recently utilized OpenAI models to identify and fix critical flaws in EU project code, while CERT Polska employed specialized GPT models to uncover vulnerabilities in router software. These efforts are complemented by private partnerships, such as the AI-based review of the Cyber Resilience Act's reporting platform by AISLE.
This shift occurs alongside the implementation of the Cyber Resilience Act, which mandates strict reporting timelines for manufacturers—including a 24-hour early warning window—under threat of significant financial penalties. While AI has accelerated the speed of analysis, technical teams maintain that human verification remains essential to validate hypotheses and judge the actual impact of discovered flaws. The overarching strategy involves building secure testing platforms to narrow the window between vulnerability disclosure and exploitation.
Full Take
The strongest version of this narrative is one of necessary evolution: as attackers weaponize flaws within minutes, defenders must adopt the same frontier AI tools to maintain parity. This is a pragmatic race for "cognitive speed" in cybersecurity.
However, a pattern emerges regarding the dependency on a narrow set of US-based providers. The narrative emphasizes the "pressure" and "months of requests" the EU faced to gain access to models like Mythos and GPT. This frames the US AI labs not just as vendors, but as gatekeepers of essential security infrastructure. While the text avoids explicit alarmism, the urgency is underscored by the "narrowing window" of defense, creating a framework where security is contingent upon the benevolence or partnership of a few private corporations.
The root cause is the asymmetry between AI capability and sovereign infrastructure. The EU is currently in a position of "trusted access" rather than "autonomous capability." This echoes historical patterns of technological dependency where the tools used to secure a system are owned by entities outside the system's legal jurisdiction.
The second-order consequence is a potential feedback loop: as AI finds flaws faster, the pressure for stricter reporting laws (like the CRA) increases, which in turn increases the demand for AI tools to manage that reporting.
Patterns detected: none
Counterstrike Scan: A coordinated campaign would likely weaponize this by framing the EU as "helpless" without US AI to provoke political instability or push specific vendor contracts. The actual content remains descriptive and does not match this attack pattern.
Bridge Questions:
1. If the tools used to find vulnerabilities are proprietary and closed-source, how can the findings be independently audited for false negatives?
2. What happens to EU cybersecurity resilience if "trusted access" to these specific US models is revoked or restricted?
3. Does the speed of AI-driven discovery fundamentally change the nature of "responsible disclosure"?