The 26th Assises de la cybersécurité took place from 7 to 9 October 2026 at the Grimaldi Forum in Monaco. Among the findings that structured the discussions, one topic came out of the shadows: data leaks have become the most frequently experienced incident, and for the first time, the most openly discussed. This shift in posture reflects an operational reality, and AI agents are its new acceleration vector.
A candid conversation about data leaks
For years, organisations that experienced a data leak avoided discussing it in professional circles. Fear of reputational damage, client questions and regulatory scrutiny pushed them towards silence. The 2026 Assises marked a turning point: the Journal du Net, which has covered the event since its inception, led with the headline that "the silence around data leaks has finally broken" (Journal du Net, 13 October 2026).
This is not an abstract cultural shift. It is the product of an incident volume high enough that many organisations attending Monaco had experienced one in the preceding eighteen months. When experience is shared, the exchange becomes useful and silence becomes counter-productive.
For your organisation, this signal has a direct implication: your partners, service providers and clients now have their own experience of data leaks. The question is no longer about the probability of an incident, but about the quality of your response when it occurs.
AI lowers the barrier to entry for attackers
At Monaco, Guillaume Vassault-Houlière, Chief Executive of YesWeHack, made a straightforward observation: "With AI, anyone can launch cyberattacks" (Assises de la cybersécurité, 8 October 2026). This statement encapsulates a shift that threat models must now incorporate.
Targeted exfiltration campaigns, which previously required organised teams working over several weeks, can now be carried out by an individual with no advanced technical skills. An AI agent can identify high-value files in a system, compress them and transmit them to an external service, while adapting its behaviour to bypass common detection heuristics. This is not an evolution in techniques; it is an evolution in the cost of access to those techniques.
The exposure volume grows mechanically. By way of illustration, approximately 3,000 CVEs were published for the Chrome browser alone in 2026, compared with fewer than 250 the previous year (figures cited at the Assises de la cybersécurité, IT for Business, 13 October 2026). Every unpatched vulnerability is a potential entry point for this type of automated attack.
The AI agent as an uncontrolled internal entity
Radware's presentation at the Assises highlighted a threat vector specific to organisations deploying AI agents internally. The phrasing used summarises the problem: an AI agent can "be seen as a legitimate entity with access to the most sensitive data" (Radware, Assises de la cybersécurité, October 2026).
Your behavioural detection tools are calibrated for unusual human actions: an account accessing abnormal volumes of files, at atypical hours, from an unenrolled device. An AI agent operating with valid tokens does not necessarily trigger these heuristics. It works at machine speed, consults thousands of documents in minutes, and appears in your access logs as an authorised application.
A concrete example: an AI agent deployed to automate the preparation of your compliance reports has, by design, read access to your security policies, risk mapping documents and supplier contracts. If this service account is compromised or misconfigured, the entire documentary asset can be exfiltrated with no behavioural alert triggered.
This is directly linked to the management of non-human identities, whose scope NIS2 and DORA programmes still underestimate. The article on non-human identities as a systemic blind spot covers the governance obligations these entities generate in your compliance programme.
What your security programme must revise
The findings from the 2026 Assises point to three operational adjustments.
Extend DLP monitoring to API flows and agents. Data loss prevention typically covers endpoints, messaging systems and file-sharing tools. Outbound flows via third-party APIs or internal AI agents often fall outside this scope. Start with an inventory of authorised data outputs, including non-human identities and AI services deployed in your environments.
Prepare your GDPR notification process for a higher volume of incidents. Article 33 of the GDPR requires notification to France's national data protection authority (CNIL) within 72 hours of establishing that a personal data breach has occurred. If the volume of incidents rises, this deadline becomes a staffing criterion for your response team. A purely manual process does not scale. Organisations that have implemented automated qualification workflows, with a human decision at the end of the chain, meet this requirement in a structured way.
Test your alerts against agent-driven exfiltration scenarios. Your exercises must include patterns specific to AI agents: high volumes of access over a short period, access to sensitive resources from a service account, transmission to an external endpoint. These patterns do not resemble a classic human attack and will not trigger your alerts unless you have explicitly configured them to do so.
Compliance is no longer a sufficient end goal
The theme of the 2026 Assises, "Osons!" (Be bold!), invited security leaders to take a strategic role within their organisations. That message applies equally to how you design your security programmes. A programme built to satisfy an annual audit is not dimensioned to detect an exfiltration carried out by an AI agent in a matter of minutes. The gap between formal compliance and genuine risk control did not narrow in 2026; it widened.
If you wish to assess your coverage against the exfiltration vectors identified at the 2026 Assises, our security tooling integration service includes a review of outbound data flows and the configuration of detections suited to agentic environments.