Janis Hecker, Inken Paland, Arian Abassi, Regina Overchyk © Impact Hub Berlin

<aside> 🔎

Summary of the panel discussion “AI Safety for Builders”, examining AI safety from three angles: the technical realities of how products fail under adversarial conditions, the longer-horizon risks that systemic deployment of AI introduces, and the regulatory framework that EU founders are now legally required to navigate.

The panel was a final event in the series of workshops “Work in AI Era”, organized by Founder2Founder and Impact Hub Berlin.

04.06.26, Berlin

</aside>

Participants:

Arian Abbasi, Responsible AI | AI Safety | Red Teaming @ Accenture

Arian Abbasi is a Responsible AI R&D Engineer based in Düsseldorf who co-founded Accenture's first multi-agentic AI red-teaming platform. He focuses on shaping scaled safety testing at AI frontiers in Silicon Valley and around the world to secure novel safety standards for major technology products.

Inken Paland, Responsible AI | AI Safety communicator

Inken Paland is a Berlin-based AI safety communicator, keynote speaker, and founder. She translates the risks and opportunities of advanced AI for mainstream audiences.

Janis Hecker, Head of AI - Regulation & Strategy @ Bitkom

Janis Hecker is Head of AI Regulation & Strategy at Bitkom e.V., Germany's leading digital industry association, covering AI regulatory policy, the EU AI Act implementation, the GPAI Code of Practice, and Germany's AI competitiveness agenda, as well as co-authoring several Bitkom publications on topics ranging from AI Act standardization to shadow AI in enterprises.

Moderated and organized by Regina Overchyk, founder of Collective Research, supporting AI Safety research in Europe.

SECTION 1 | 💻 The everyday risks of building with AI

Speed and automation create real liability — and founders are closer to the consequences than they think.

Moving fast with AI-assisted development introduces failure modes that weren't common in traditional software. Database errors from unreviewed code pushed on a Friday, weekend incidents from Claude Code or similar tools running without sufficient checks, and unexpected user behaviour are all increasingly routine.

A recurring theme was misuse: users treating AI-powered products in ways their creators never intended. One illustrative example — a car dealership's sales bot being used by visitors to complete their homework — is relatively harmless. Others are not. Founders should assume that some percentage of users will interact with their product around sensitive topics including self-harm.

Key points from this part of the discussion:

• Adversarial testing should be built in from the start, not added after launch. Ask someone outside your team to try to break the product — younger people tend to be especially effective at this.
• At scale, manual review is impossible. LLM-as-a-judge — running a separate model over all your traces — can flag outputs that fall outside your intended policy.
• Formalise your policy. What counts as an out-of-bounds output? Large companies have teams writing this down. Small teams can approximate this with structured prompts and automated checks.