Piktogramm mit Kopfsilhouette und Sprechblase auf violettem Hintergrund

Vulnerabilities in AI Systems

The State of IT Security in Germany in 2026

Just over half of the vulnerabilities that became known in agentic ecosystems had a high or critical severity (cf. Figure 2-6).

Figure 2-6: Newly Disclosed Vulnerabilities in AI Agents by Severity

Unit:

number

Note: CVEs relating to AI agent systems. AI agent systems include AI coding assistants, general AI agent systems and enterprise agent assistants. This does not include agent frameworks, workflow and orchestration platforms, or tool gateways. This product uses the NVD API but is neither recommended nor certified by the NVD.

The majority of the vulnerabilities did not affect the AI model itself. Rather, the interfaces through which the agents access other tools, read files, execute code, or interact with APIs and databases were affected. The secure design of these interfaces – in particular the securing of tool calls, access to external services, and authorization concepts – is a central leverage point for the resilient design of AI ecosystems.

The impact that individual products have on the statistics is shown by the agent system OpenClaw: for this one alone, 539 vulnerabilities were recorded in the reporting period (cf. Selected Incidents: Attack Surface - Vulnerabilities in AI Agent System OpenClaw).

In addition, 99 new vulnerabilities that enable prompt injection attacks became known in the current reporting period (cf. Figure 2-7).

Figure 2-7: Prompt Injection Vulnerabilities in AI Systems by Severity

Unit:

number

Note: CVEs whose exploit chain contains prompt injection as a mechanism. Without vulnerabilities of unknown severity. This product uses the NVD API but is neither recommended nor certified by the NVD.

Attackers induce an AI system to unintended behavior through manipulated inputs. Around 60 of the prompt injection vulnerabilities that became known and were analyzed in the current reporting period enable attackers to induce the AI system to execute code – including malicious code. The impact can thus go far beyond mere falsification of result outputs, even though attackers do not need to install malware for this. By means of so-called indirect prompt injections, they instead embed harmful instructions directly into content that an AI system processes automatically (such as e-mails, websites, or documents). In technical terms, the resulting attack chain – from the initial prompt injection through privilege escalation to nesting – is also referred to as promptware.

It is to be assumed that AI-specific vulnerabilities, which, for example, enable the poisoning, i.e., manipulation, of training data, or the targeted deception of AI models, are still underrepresented in the vulnerability statistics. Securing multifunctional AI agent systems is extremely complex, and due to the dynamic development, established security standards are still lacking. Classic IT security measures such as input validation, output checking, and access restriction, however, retain their relevance.