What Check Point reports
The September report is part of a monthly series. Two groups of figures come from it. First, the attack picture: worldwide an average of 2,803 attacks per organization per week, 48 percent above the same month last year and 16 percent above August. Second, the data employees enter into GenAI tools: in September, per the report, one in every 39 prompts in corporate networks carried a high risk of data leakage, at 89 percent of regularly using organizations.
The German figures come from a media report on the Check Point report: 2,316 attacks per organization per week, 103 percent more than in September 2025. Check Point itself mentions Germany in its blog only in the ransomware section, with a 4.0 percent share of reported victims. We therefore couldn't read the German attack figures in the original blog.
The context missing from the headline
'More than doubled' sounds like a German exception. In the report, though, Germany's value sits below the global average and below Austria's. The increase isn't regionally unique either: Austria was up 96, Switzerland 99 percent, Europe as a whole 61 percent. The source also doesn't say how an 'attack' is counted or what telemetry the figures rest on. They work as an order of magnitude, not as a measurement.
The prompt rate is similar. Europe has the lowest rate of any region at 1 in 57. And the rate moves from month to month: 1 in 36 in July, 1 in 43 in August, 1 in 39 in September. What follows is that the September value is nothing special, not that the problem is growing or shrinking. '89 percent of organizations affected' also means at its core: almost every company that regularly uses GenAI has at least one risky prompt a month. That's less a statement about severity than about prevalence.
What the data types show
More revealing for companies is which kind of data lands in prompts. Per Check Point, network and IT infrastructure data appeared in 71 percent of organizations, financial data in 70 percent, legal and regulatory data in 68 percent, employee and HR data in 62 percent, and personal data in 60 percent. Important: the report stresses these percentages mean the share of organizations where the category was observed, not the share of prompts.
For German companies, HR and personal data are particularly relevant because they directly raise GDPR questions: anyone entering personnel or customer data into an external AI tool needs a legal basis and, as a rule, a data processing agreement. That's our reading, not a statement in the report. It doesn't say whether the captured cases were legally problematic, only that the data types occur.
By industry, business services lead at 4.9 percent (about 1 in 20 prompts), ahead of financial services at 4.1 and healthcare at 3.5 percent. For consultancies, agencies, and other service providers in our audience, that's a hint to review your own practice, but no evidence of your own rate.
What the report doesn't settle
The blog doesn't define 'high risk'. It names neither sample nor customer base nor how many organizations were measured. That's usual for vendor reports and no accusation, but it means: the figures are a snapshot of the data of a security vendor selling protection against exactly this risk. We aren't saying they're wrong. We're saying they can't be independently verified.
What this means in practice
- With vendor security figures, separate increase from absolute level and check how 'attack' or 'high risk' is defined.
- Take your own inventory of which GenAI tools are actually used in-house. Eight per organization on average is more than many expect.
- Decide which data types don't belong in external tools (personnel, customer, financial, legal data) and where exceptions with a contract are possible.
- Pair that with a technical control, not just a policy. A rule nobody sees has little effect.
- Don't monitor employees' prompts across the board without involving the works council and data protection. Monitoring has its own legal limits.
- Don't sell or read single monthly values as a trend.
The core: the figures are plausible as an order of magnitude and need context. More important than any monthly value is the question of which data lands in your prompts and whether you know.