An OpenAI model found an 8.8 token-replay flaw in code for an EU project

The 8.8 is a missing state check. According to the OpenCVE record Politico tied to the finding, the software never registered that an activation or recovery link had been used. An attacker holding one live link could replay it while it was valid and reset the password again ^[5][6]. The affected code is Vulnerability-Lookup, an open-source tool for tracking vulnerabilities, in versions up to 5.5.1 ^[4]. There is no crash in it and no memory corruption. The failure is a rule the code never enforced. TNW, which summarised the Politico report, said it had not independently verified it ^[8].

CERT Polska published six MikroTik RouterOS flaws on 5 September. The team said the models sped up the analysis but that it still tested every hypothesis on real systems, and that its own staff judged the impact of each flaw ^[9][11]. Two of those six are already being chained by attackers in a combination the team calls MikroTrick, which gives full control of routers that expose SSH to the internet. CERT Polska has seen the attacks since at least 2 September ^[12][13]. The advisory came three days later ^[14].

Across the two teams the published count is ten flaws found with an OpenAI model, four from ENISA and CERT-EU and six from CERT Polska ^[15]. Two of the ten are under active attack ^[12][15].

OpenAI's head of policy for Europe, Tom Duff Gordon, spoke of a "narrowing window" for AI to find weaknesses before attackers do, in a statement quoted by Politico ^[16]. "That's why we work with partners such as ENISA," he said ^[17]. ENISA is in OpenAI's trusted access programme for cyber models, which the company runs through its Daybreak cyber defence programme ^[18]. Heuvinck told Politico that ENISA and CERT-EU have used advanced models since July to "actively scan" EU institutions, and under a Commission action plan published the same month the agency and the Joint Research Centre are building a "secure testing platform" for the most advanced models ^[31][32].

The second review is AISLE's. AISLE said on 14 September that it had reviewed the code of the Single Reporting Platform for the Cyber Resilience Act and will keep reviewing it for 12 months ^[22][23]. Manufacturers of products with digital elements sold in the EU must now use that platform to report exploited flaws and severe incidents ^[25]. The Register reported the deadlines: an early warning within 24 hours, a fuller notice within 72, and a final report within 14 days of a fix, with the 24-hour deadline applying wherever the manufacturer is based ^[26]. Breaching those duties can bring the act's top fines of 15m euros or 2.5% of annual turnover, whichever is higher. Most of the act's other rules apply from 11 December 2027 ^[27][28]. ENISA's Hans de Vries said the agency had also run user and security testing with stakeholders including national CSIRTs. He added: "I thank AISLE for their important AI-based secure code review performed" ^[29][30]. The report does not say what the AISLE review found.

For whoever has to defend this to a security lead, the question is what has to be true before a model-flagged finding leaves the tracker: reproduction on a real system, and a named person who signs the severity. CERT Polska published both of those alongside its model names ^[11]. ENISA's four are fixed, and the 8.8 has a record dated 12 August behind it ^[2][7].