Data study ·
AI Security Report 2026: Check Point reveals how AI has gone from accomplice to mastermind of cyberattacks
Original title, in French: “AI Security Report 2026 : Check Point révèle comment l’IA est passée de complice à cerveau des cyberattaques”
- Publication
- IT for Business (France)
- Author
- Laurent Delattre
- Published on
- 14 October 2026 · read in the page structured data
Summary
Check Point Research's "AI Security Report 2026" describes an AI that now directs cyberattacks rather than simply speeding them up. Examples include VoidLink, a sophisticated attack tool created by a single developer assisted by a coding AI, and the breach of nine Mexican public bodies by a lone hacker using Claude Code and GPT-4.1, exposing 400 million records. The report also points to an internal risk: companies are deploying AI agents on a large scale, often without IT department oversight, creating a new attack surface. Adrien Merveille, Chief Technology Officer of Check Point France, cites figures on MCP server vulnerabilities and high-risk prompts, and calls for AI agents to be governed like employees, with defined permissions.
Summary based on the full text of the article.
Read the article on IT for Business →Figures cited
-
5 x
Increase in detections of long malicious prompt injection payloads between March and May 2026
March–May 2026, Check Point telemetry
« Les détections de longues charges malveillantes d’injection de prompts ont été multipliées par cinq environ entre mars et mai 2026. »
Source: AI Security Report 2026, Check Point Research
-
40 %
Share of reviewed MCP servers showing weaknesses
10,000 MCP servers reviewed by Check Point
« Check Point a relevé des faiblesses sur 40 % des 10 000 serveurs MCP (Model Context Protocol) examinés »
Source: AI Security Report 2026, Check Point Research
-
70 %
Share of OpenClaw agent "skills" requesting more credentials than necessary
skills analysed for the OpenClaw agent
« 70 % des « skills » analysées pour l’agent OpenClaw réclamaient plus d’identifiants que nécessaire »
Source: AI Security Report 2026, Check Point Research
-
10 applications
Average number of AI applications used per organisation per month
organisations, per month
« En moyenne, une organisation utilise dix applications d’IA par mois, souvent sans le feu vert de la DSI et hors de tout processus d’approbation. »
Source: AI Security Report 2026, Check Point Research
-
4 %
Share of high-risk prompts (sending sensitive data to external services)
in 2026, compared with 2% a year earlier
« La part des prompts à haut risque, ceux qui envoient des données sensibles ou réglementées vers des services externes, a doublé en un an, de 2 % à 4 % »
Source: AI Security Report 2026, Check Point Research
-
90 %
Share of organisations recording at least one high-risk prompt interaction per month
between 87% and 93% of organisations, per month
« entre 87 % et 93 % des organisations enregistrent chaque mois au moins une interaction de ce type »
Source: AI Security Report 2026, Check Point Research
-
3.95 %
High-risk prompt rate in Europe
Europe
« L’Europe affiche même le taux le plus élevé de toutes les régions : 3,95 % contre 3,45 % en moyenne mondiale. »
Source: AI Security Report 2026, Check Point Research
-
3.45 %
Global average high-risk prompt rate
global average
« L’Europe affiche même le taux le plus élevé de toutes les régions : 3,95 % contre 3,45 % en moyenne mondiale. »
Source: AI Security Report 2026, Check Point Research
-
88,000 lines of code
Lines of code produced by a single developer using an AI tool for VoidLink
development of VoidLink, in under a week
« 88 000 lignes de code fonctionnel en moins d’une semaine. »
Source: AI Security Report 2026, Check Point Research
-
400,000,000 records
Number of records exposed in the breach of Mexican public bodies
nine Mexican public bodies, December 2025–February 2026
« un pirate isolé, qui n’a toujours pas été identifié, a infiltré neuf organismes publics mexicains, dont l’administration fiscale, et exposé quelque 400 millions de fiches »
Source: AI Security Report 2026, Check Point Research
-
5,317 commands
Number of commands executed by the AIs during the attack in Mexico
34 sessions, December 2025–February 2026
« En 34 sessions, le pirate a donné 1 088 consignes, et les IA ont exécuté d’elles-mêmes 5 317 commandes. »
Source: AI Security Report 2026, Check Point Research
Classification
- Country
- Mexico
- Company size
- All sizes
- Industries
- Software vendors, IT services companies, telecoms · All industries, or industry not specified
- Companies
- Check Point (study author), OpenAI (cited), Anthropic (cited)
- Keywords
- cybersecurity, AI agents, shadow IT, prompt injection, data security, agentic AI
Entry written by Marsanne Studio from the cited article, with the help of an AI model (claude-sonnet-5). The rights to the article and its illustrations belong to IT for Business and its author. Entry translated from French by Claude Sonnet 5.5.