"transparency," said Thomas Regnier, welcoming OpenAI's move and saying it will allow the EU executive body to "follow the deployment of this model very closely, but also to potentially address certain security concerns in a closer way."
OpenAI's Trusted Access for Cyber program and GPT-5.5-Cyber
OpenAI has agreed to provide European authorities and companies with restricted access to its vulnerability-finding model, GPT-5.5-Cyber, through a new Trusted Access for Cyber program. The company said dozens of European organizations will be allowed into the program to "play with restricted access models" that can find and exploit vulnerabilities — capabilities the company frames as tools for defenders to protect systems and respond to threats more quickly.
OpenAI EMEA managing director Emmanuel Marill said the firm must "block dangerous activity, while making sure trusted defenders have tools that are genuinely useful in protecting systems, finding vulnerabilities and responding to threats quickly." OpenAI also noted that the head of its OpenAI for Countries initiative, former U.K. chancellor George Osborne, presented an action plan to the commission and EU member states that included access to the latest models and briefings with OpenAI's safety and security teams.
Which EU bodies will see the model — and the next steps
At a press briefing the European Commission confirmed OpenAI had agreed to let it "poke around" GPT-5.5-Cyber, but said details remain to be settled. Commission spokesman Thomas Regnier said talks will continue this week, and one major outstanding question is which commission entity will get access. Candidates named by the commission include the commission's new AI Office, its directorate-general for technology (which employs a cybersecurity director), and the EU cybersecurity agency ENISA. "One step at a time," Regnier added.
Regnier declined to draw implications from the differing stances of other AI vendors, saying only that "with one [AI firm] you have a company proactively offering to give access… with the other one we have good exchanges [and] we're not at a stage where we can speculate on potential access or not."
Named participants and industry partners
OpenAI said participating organizations will include Deutsche Telekom, Telefonica, Spanish bank BBVA — which is also becoming a founding partner in the new OpenAI Deployment Company — and German digital investment platform Scalable Capital. U.K.-based cybersecurity firm Sophos will also join; Sophos chief technology officer John Peterson said the company will be "leveraging these capabilities" in Sophos MDR, its agentic security operations center.
The program aligns with similar restricted-access efforts in the sector; the report contrasts OpenAI's approach with Anthropic's, noting Anthropic has withheld comparable access and operates a separate initiative called Project Glasswing. OpenAI did not respond to a request for comment on why it is willing to give European organizations access while Anthropic is not.
Bafin's warning: patching cycles, a new division, and the incident tally
Germany's financial regulator Bafin used the news cycle to underline urgency. In a speech marking the release of its latest annual report, Bafin president Mark Branson warned that frontier AI models will increase pressure on cybersecurity and will require faster patching: measures "good today will be far from sufficient tomorrow." He said patch management cycles "in the past" could be measured in months, but going forward companies will have to complete them "within a few days, if not hours," a change he called especially challenging for small and medium-sized enterprises.
Branson also announced the creation of a new division inside Bafin's Directorate for Cyber Risks and Technology in the Financial Sector to conduct speedy and targeted "IT spotlight" inspections. He said inspections regularly uncover problems, "particularly with patch management," and that more targeted inspections will help both the regulator and companies better assess risks and strengthen resilience.
Bafin, which became Germany's central reporting hub for serious ICT incidents with the EU Digital Operational Resilience Act (DORA) last year, recorded 733 incident reports over 2025. About a tenth of those related to cybersecurity incidents, mostly cyberattacks: 31.1% phishing, 28.3% malware and hacking, 15% denial-of-service, and 13.3% ransomware. Two-fifths of the attacks affected service providers rather than the financial entities themselves. Bafin warned of rising dependency on outsourced service providers, often based outside the EU, and noted concerns that foreign authorities might be able to demand access to European companies' data.
What this means for cybersecurity teams, EU policymakers, and financial institutions
- Cybersecurity teams and security operations centers: The Trusted Access for Cyber program promises powerful vulnerability-discovery tools that defenders can use; Sophos said it will leverage these capabilities in Sophos MDR. At the same time, Bafin's warning raises an operational imperative: patching cycles must accelerate from months to days or hours.
- EU policymakers and regulators: The commission must decide which internal body — the AI Office, the directorate-general for technology, or ENISA — will receive access to GPT-5.5-Cyber, and continue negotiations about oversight and deployment. OpenAI's offer creates an opportunity for the commission to "follow the deployment" of a frontier model more closely, per Thomas Regnier.
- Financial institutions and outsourced service providers: Bafin's incident data and new inspection division signal closer regulatory scrutiny. The regulator flagged that two-fifths of attacks hit service providers, and it warned of data-security doubts when providers are headquartered outside the EU.
OpenAI's move to open restricted access to GPT-5.5-Cyber for European actors sets up a test of two linked bets: that giving trusted defenders powerful tools will measurably improve resilience, and that regulators can keep pace with deployment decisions on models that find and exploit vulnerabilities. The immediate next steps are procedural — which EU body will see the model, and the outcome of the commission's talks this week — while on the ground Bafin has signaled regulators and companies must prepare to patch faster and face more targeted inspections.
