A central component of the action plan is the Bavarian AI Security Dialogue. In this new format, representatives from industry, science, and politics will regularly assess current developments in AI security and derive necessary actions.

The dialogue is to take place at least twice a year and can also convene at short notice in the event of particular developments. The first session is still scheduled for 2026.

The results of the preceding AI summit are to be documented in a detailed outcome paper. This will serve as a technical basis for the first meeting and be coordinated with the Bavarian AI Council.

Among those involved in the summit was Prof. Dr. Sonja Zillner, member of the Bavarian AI Council. Other participants came from companies such as Google, IBM, Mistral AI, and OpenAI, as well as from academia, ethics, and society.

Annual report to create an independent situation assessment

The security dialogue is to be accompanied by independent scientific support commissioned by the Bavarian Ministry of Digital Affairs. The person responsible for this task is to be determined as early as 2026.

Based on this, the ministry plans an annually published Bavarian AI Security Report. The first report is to be published in 2027. The goal is to build its own scientific expertise and a reliable situation assessment of the safety of powerful AI systems.

The background is particularly the growing capabilities of frontier models and agentic AI. The ministry refers to safety tests in which powerful models have shown problematic behavior such as deception or bypassing safeguards under constructed test conditions.

Own AI ethics forum planned

In addition to technical security issues, Bavaria wants to institutionally accompany the social and ethical impacts of powerful AI more closely. For this purpose, a Bavarian AI Ethics Forum will be established under the umbrella of the Ministry of Digital Affairs. This body is also to convene for the first time by 2026.

In terms of regulation, the ministry advocates a risk-based approach. The requirements are to be oriented to the capabilities of an AI system and its potential for harm. Highly autonomous frontier models are to be treated differently from limited assistance systems.

AI security is also to be related more closely to cybersecurity . Apart from the risks of autonomous systems, it is also about the targeted misuse of powerful AI for cyberattacks, fraud, manipulation, or disinformation.

Bavaria again advocates for a national AI security institute

In parallel, Bavaria renews its call for a German AI Security Institute and advocates for it to be located in the Free State. The Ministry of Digital Affairs intends to address this demand again to the federal government.

Moreover, the results of the Munich summit are to be discussed at the European level. Mehring intends to present them at the next Digital Ministers' Conference in Brussels in November.

Additionally, the ministry plans new information and dialogue offerings to inform citizens about the functioning, possibilities, and risks of AI. Both traditional and social media channels are to be used for this purpose.