AI tool use unregulated in many companies

A study by WatchGuard Technologies in April found that 64 percent of employees in small and medium-sized enterprises and mid-sized companies across eight countries, including France, use artificial intelligence tools not validated by their management. Less than 30 percent of the organizations surveyed maintain an inventory of these tools.

Lack of clear guidelines for AI use

The study also revealed that nearly four out of ten companies do not know which AI applications their teams actually use. According to a 2025 Salesforce study in France, 58 percent of professional users of generative AI work without a charter or framework defined by their employer. At Impact AI, a group of French companies in the sector, only 9 percent of employees had an ethical charter or a dedicated contact person that year.

AI tool restrictions vary globally

Another study by Okta found that France is the country where the phenomenon of bypassing AI tool restrictions is the least widespread, far below the global average of 52 percent. This suggests that French organizations may be more compliant with internal AI policies compared to others.

Security concerns with AI tools

The CNIL published new recommendations on generative artificial intelligence last year, urging companies to inform individuals when their personal data may be memorized by AI models and to provide means for individuals to exercise their rights of access and correction, even when employees use personal accounts. The ANSSI recorded 42 groups of hackers exploiting AI to prepare attacks in February of last year.

Malicious AI tools in underground markets

North Korean groups, such as Lazarus, create fake professional profiles on social networks to approach targets. Distracted versions of AI tools, such as WormGPT or FraudGPT, circulate on underground forums without security limits for the equivalent of a hundred euros per month.

Vulnerabilities in AI assistants

In 2025, Aim Security discovered a vulnerability in Microsoft 365 Copilot called EchoLeak, which allowed an attacker to send a trapped email that could make the AI assistant recite the content of messages and professional files it had access to. Microsoft fixed this vulnerability, but seven months later, another incident was reported in Copilot Chat, integrated into Outlook, where confidential emails in the Sent and Drafts folders were accessible to the assistant due to a coding error.

Risks of fake software packages

Security researchers identified a mechanism called slopsquatting, where a code assistant might recommend the installation of a non-existent software package. Between August 4 and 6, security researcher Paul McCarty recorded nearly 800 such packages on the npm manager. A fake module imitated a legitimate mobile payment service under the name "checkout-mobile-bnpl," triggering malware adapted to the victim's operating system.

Recommendations for securing AI systems

The ANSSI guide compiles 35 recommendations to secure generative AI systems. It advises companies to create an internal charter specifying which AI tools employees can use and with what data, have each new employee sign it, maintain an inventory of AI tools used, limit access permissions for AI assistants, log their activity, review all code produced by AI assistants before deployment, verify package names before installation, avoid automating critical actions without human validation, prefer qualified hosts for internal AI tools, conduct training sessions before introducing new tools, and inform employees of their rights over personal data entrusted to AI tools. The CNIL has required this since February 7, 2025.