In a German Survey, 64% of AI-Using Firms Limit Access Rights and 40% Log Use
1Nearly Two in Three AI-Using Firms Limit Access Rights and Two in Five Log Use
Of the 246 German companies that currently use AI applications, 64% restrict access rights and 40% log or monitor use, while 8% have taken none of the technical safeguards the survey asked about. Germany's Federal Office for Information Security (BSI) and the TÜV-Verband, the association of the TÜV technical inspection and certification organizations, published the survey on AI and cybersecurity in companies on October 7, 2026. The research firm forsa phoned 505 German companies between June 22 and July 31, 2026. The survey covers company AI use in general and gives autonomous agents a short section of their own. Those two shares show how widely companies restrict AI access and keep a trace of it.
Companies are introducing AI faster than they adapt their security structures. Only 11 percent of those who use or plan to introduce AI have their own processes for AI-related security incidents. Risk assessments, restricted access rights, tests, logging and control are far from established everywhere.
Dirk Stenkamp, president of the TÜV-Verband and CEO of TÜV NORD, in the study's foreword (translated from German)
Among AI-using companies with 250 or more employees, 80% restrict access rights. The question records whether a safeguard exists, in a multiple-answer list, and the chart does not show how far each one extends.
2Most Companies See Agents as a Risk Even Though Few Use Them
Only 5% of companies that use AI run autonomous agents today, and another 8% plan to. The survey asked all 505 companies about agent risk anyway, describing agents as systems that plan and run multistep tasks on their own, for example with access to internal ticketing systems, databases or email. 84% rated wrong decisions based on false or invented findings a very or rather high risk. 78% gave three more risks the same rating: new security risks through interfaces to other systems, leaks of sensitive data or business secrets and the AI triggering unwanted actions. 74% gave the same rating to unclear responsibility for errors. A report published this week on coding agents posting internal screenshots to public GitHub repositories describes agents working on access each developer already held.
When deploying AI agents, special attention should go to technically limited rights, controlled interfaces and suitable test environments. Critical or irreversible actions should be secured proactively through human approvals, transaction limits and technical abort options.
BSI and TÜV-Verband, recommendations section (translated from German)
The recommendation sits in a section on independent testing and certification. The study's foreword says the TÜV organizations are developing three AI certification programs, and a pilot is running with partners, with availability planned for the end of 2026.
Source: BSI and TÜV-Verband, KI-Sicherheitsstudie 2026, October 7, 2026.