Artificial Intelligence is becoming an active cyber attacker, moving from just assisting attacks to building malware, supporting intrusion operations, accelerating vulnerability discovery and participating in post-exploitation activities, according to a report by CheckPoint Research.
“AI is increasingly enabling attackers to execute operations faster and at greater scale, significantly compressing the time between vulnerability disclosure and exploitation.” It added that Anthropic’s Claude Code now carries out 80–90% of the tactical work across approximately 30 target organisations. This highlights that finding vulnerabilities is becoming more automated, requiring fast action.
AI adoption has expanded markedly in recent years. According to the data by Check Point, the average number of prompts per user grew 25% to 70 in May 2026 from 56 in December 2025 on an individual basis. Highlighting the scale of risk, 87–93% of organisations reported at least one high-risk GenAI interaction each month.
The proportion of high‑risk prompts—those containing sensitive corporate, personal, or regulated data sent to external AI services—doubled, rising from 2% to 4%. This increases the baseline risk of data leakage.
Europe and Latin America recorded rates of high-risk GenAI prompts above the global average (3.45%) at 3.95% and 3.76%, respectively. This indicates that organisations in these regions face a heightened risk of sensitive data leakage. Looking at the industry, business services, wholesale & distribution, and telecommunications ran the highest risks.
In a case study, the cyber security firm said that Anthropic’s Project Glasswing uses an unreleased frontier model, Claude Mythos Preview, to find and fix vulnerabilities in critical software.
“In its first month, it autonomously identified more than 10,000 high- and critical-severity zero-days across every major operating system and browser, and successfully produced a working exploit on the first attempt in roughly 83% of cases.”
Zero-day is a vulnerability where the vendor has had “zero days” to fix the problem. Anthropic has refrained from publicly releasing the model because of concerns that it could be exploited.
