On October 21, two critical vulnerabilities were disclosed in PHP, one of the most widely used programming languages on the web. The flaws had gone undetected for more than a decade before being uncovered with the help of AI models developed by OpenAI and Google. They could allow hackers to take control of servers remotely, making this one of the most significant cybersecurity incidents in recent years. Several governments, including France and the United Kingdom, temporarily took official websites offline as a precaution. No successful breaches have been publicly confirmed, but attempts to exploit the vulnerabilities were detected as early as October 17.
That same day, Microsoft revealed that a hacking group it assessed to be linked to the Chinese government had used AI agents to automate attacks against hundreds of organizations. In several cases, the attackers gained privileged access in under three hours.
“This time, AI helped researchers find the vulnerabilities before hackers could. But it is also giving attackers the ability to discover and exploit security flaws at unprecedented speed. We cannot wait for an attack to cripple hospitals or bring down a power grid before taking action. France must commit substantial resources to protecting its critical infrastructure and secure an international agreement restricting the deployment of AI models with the most dangerous offensive capabilities.”
Cyberattacks are becoming more dangerous
These incidents are part of a rapidly growing wave of AI-assisted cyberattacks. According to a Reuters investigation published on October 19, a hacker used AI agents to compromise the websites of more than 100 companies. In Latin America, a criminal group reportedly targeted more than 400 financial institutions using AI agents powered in part by Claude and Grok.
The consequences can extend far beyond the digital world. On October 11 and 12, hackers seized control of refrigeration systems at hundreds of French supermarkets. In late 2025, a cyberattack damaged electricity generation equipment in Poland and nearly triggered a power outage. AI now threatens to accelerate attacks of this kind, while many computer systems remain outdated or inadequately protected. The UK government said the incident underscored the importance of its new public preparedness campaign, “Be Ready,” which advises households to keep a three-day supply of bottled water and non-perishable food.
CeSIA calls for four urgent measures
- Accelerate France’s cyberdefense efforts. The government must fast-track the implementation of the cyberdefense plan announced in April, provide ANSSI with additional staff, funding, and technical resources, and swiftly pass the Resilience Bill, which would implement the EU’s NIS2 Directive into French law, already two years behind schedule.
- Get ahead of emerging vulnerabilities. ANSSI and its European counterparts must be given early access to the most capable AI models for cybersecurity research. AI developers should also be required to confidentially report critical vulnerabilities discovered using their models before any known exploitation occurs.
- Strengthen European oversight. The European AI Office must enforce the risk assessment and mitigation requirements of the EU AI Act, including for open-source models that pose systemic risks. Models with critical offensive capabilities must undergo independent evaluations and face deployment restrictions wherever their risks cannot be adequately controlled.
- Negotiate a binding international agreement. France must work with its international partners to establish enforceable rules governing the development and release of AI models that pose the greatest cybersecurity risks.
A threat that has been building for months
As early as April 8, 2026, following Anthropic’s unveiling of Claude Mythos, CeSIA called for an emergency European cyberdefense plan. Six months later, incidents are mounting, while AI companies report that their models are finding security vulnerabilities at an accelerating pace.
In an open letter published in October, nearly 500 researchers, including Yoshua Bengio and Geoffrey Hinton, call for strict containment measures for AI models capable of conducting cyber operations, including keeping them disconnected from the Internet by default.