Why do chemical, pharmaceutical and process companies in Berlin need a dedicated AI security & compliance strategy?

ISO 27001 and TISAX provide frameworks for information security that can in principle be applied to AI systems, but they require specific adaptation. For AI systems, in addition to classic IT assets, models, training data and inference pipelines must be considered protectable assets. This means risk assessments, control objectives and policies must be extended to cover model versions, datasets and access paths.

Practically, the work begins with an asset inventory: which models are running in the environment? Which data sources are used? Which interfaces exist to production or lab systems? This overview is the basis for implementing technical controls such as network segmentation, encryption and IAM, as well as for organizational measures like role and responsibility definitions.

For TISAX-relevant supply chain processes, e.g. collaboration with OEMs or suppliers, auditability is central. That means audit logs, traceability of changes to models and clear change management processes must be implemented. Automated compliance reports help provide repeatable evidence and pass audits efficiently.

Our recommendation is a pragmatic, stepwise approach: first secure the critical paths (e.g. models that influence production processes), then build up governance and documentation processes and finally integrate the remaining assets. This way companies quickly achieve an auditable baseline and can expand the depth and breadth of their security measures from there.

The decision between self-hosting and cloud operation is not only technical but also heavily dependent on regulatory, legal and organizational factors. For many pharma and chemical applications involving sensitive data or IP, self-hosting is often the preferred choice because it offers maximum control over data, access and infrastructure. Self-hosting simplifies compliance requirements such as data locality, specific retention rules and controlled access rights.

Cloud providers, on the other hand, offer scalability, managed services and often built-in security features. In many cases a hybrid approach makes sense: critical models and sensitive datasets on-premise, less sensitive workloads or development environments in certified clouds. It is important that cloud deployments are accompanied by clearly defined data-processing agreements, encrypted transfer paths and controlled access.

Technically, self-hosting must ensure that infrastructure components like HSMs for key management, secure network separation and regular security updates are in place. For cloud solutions IAM, VPC configurations, KMS and audit logging are crucial. Ultimately the choice also depends on audit requirements and internal policies.

We recommend a risk analysis that considers legal requirements, business risks and operational capabilities. For companies in Berlin with access to cloud talent, hybrid operation can offer the best balance of control and agility.

A PIA for an AI project starts with a clear description of the project: which data is used, which processing steps are planned, what outputs are generated and who has access? In lab projects personal data, genealogical or patient-related information and proprietary measurement data are often involved — each of these data types requires specific protective measures.

The next step is to analyze risks: re-identification, faulty inference, data leaks and misuse potentials. For each risk technical and organizational countermeasures should be defined, for example pseudonymization, access restrictions, output filters and logging. These measures are assessed and prioritized against the remaining residual risks.

The PIA must also document how the principle of data minimization is implemented, which legal basis supports the processing and how data subject rights are fulfilled. For research projects additional requirements such as ethics approvals or consents may be relevant — close coordination with data protection officers and, if necessary, ethics committees is essential.

Finally, a monitoring plan is recommended: the PIA is not a one-off document but part of a continuous process. Changes to the model, new datasets or altered usage scenarios require re-evaluations. This keeps the PIA relevant and protects long-term against compliance gaps.

Red-teaming is a practical testing method that challenges AI systems with realistic attack and misuse scenarios. In the process industry this is not just about data theft but about manipulation of inputs, targeted falsification of sensor values or triggering false control commands through adversarial inputs. Red-teaming uncovers such vulnerabilities before they can cause damage in production.

A structured evaluation includes black-box tests (external attacks), white-box analyses (code and architecture reviews) and scenario tests that simulate human operators and real process conditions. Clear test criteria, acceptable error rates and emergency processes must be defined, which are also integrated into BCP/DR plans.

Red-teaming should be conducted regularly and be part of the model release process. Results must be translated into concrete measures: hardening the model API, additional input validations, output sanitization or adjustments to access controls. Without this feedback loop red-teaming remains an exercise rather than an operational protective measure.

For companies in Berlin it makes sense to involve external red-teamers with domain-specific experience and to leverage local security communities to stay up to date with the threat landscape. The combination of internal tests and external review provides the best protection.

A well-focused proof-of-concept (PoC) can be realized in a few weeks to a few months, depending on scope, data availability and integration complexity. At Reruption a typical PoC starts with use-case definition, a feasibility check, rapid prototyping and clear success metrics — just like in our standardized AI PoC offering.

For AI security & compliance an MVP often means: secure self-hosting for a model, implemented audit logs, initial data governance elements (classification, retention) and a simple evaluation/red-teaming round. These core building blocks enable initial audit-readiness and demonstrate operational improvements.

Preparation is important: available datasets, clear interfaces and decision authority in the project team significantly speed up implementation. In Berlin projects benefit from easy access to tech talent and short decision cycles when stakeholders are involved early.

Realistic timelines vary: 4–8 weeks for a minimal evaluated prototype; 3–6 months for an audit-capable MVP implementation. Transparent milestones and a clear production plan ensure a PoC does not remain in the lab but makes its way into operations.

A multidisciplinary team is essential: data scientists and ML engineers develop models, security architects design secure infrastructures, compliance and legal experts assess regulatory risks and data engineers ensure clean, traceable data pipelines. Additionally, domain experts from production or the lab are indispensable to define risks and meaningful acceptance criteria.

Operational roles such as a model owner, a data quality officer and an incident response team ensure ongoing operations. For audit-readiness roles are also needed to manage documentation, change management and regular reviews. Good communication between these roles is often the key to success.

In Berlin it is often possible to quickly bring in external expertise, for example for red-teaming or specialized privacy questions. Nevertheless, critical roles should be anchored internally so that knowledge, responsibility and rapid response capability remain.

Our experience shows: teams with clearly defined responsibilities, regular training and integrated security and compliance checks reach stable, trustworthy AI systems significantly faster.

Reruption works according to the co-preneur approach: we do not act as traditional consultants but as embedded developers and product partners. Although our headquarters are in Stuttgart, we regularly travel to Berlin and work on-site with your teams to quickly clarify requirements, test prototypes and practically implement security measures.

The process starts with use-case definition and a technical feasibility check, followed by a rapid prototype (PoC) and a concrete evaluation phase. In parallel we develop a production plan with architecture, budget and timeline estimates as well as a compliance roadmap for ISO/TISAX/data protection.

In Berlin we place great value on direct collaboration with local stakeholders — from lab managers to IT security officers — to ensure implementations are practical and auditable. Our work is results-oriented: we deliver working prototypes and actionable plans, not just documentation.

If you wish, we start with a standardized AI PoC that delivers a technical validation within a few weeks. The project then scales in clear iterations toward production and audit-readiness.