Cologne AI and Machine Learning Meetup

CAIML #43 is going to happen on July 21, 2026, at IQMatix.

We will have two talks with additional time for networking.

Talk 1: Gabriel Firmino Barjollo (Cybersecurity Specialist, Pentester & AI Engineer) and Sebastian Kröger (Architect & AI Engineer): Background of jailbreaking LLMs, Pentest Agents, and Supply Chain Attacks

Modern LLM safety alignment is a moving target - and understanding where it breaks is essential to building better defenses. We introduce in the background of jailbreak techniques across prompt, system, and model levels, then demonstrate how an unaligned model becomes the reasoning core of an autonomous pentest agent, wired to standard offensive tooling with a plan-execute-observe-adapt loop. Once an AI agent acts on a developer's behalf, how far can a compromise propagate? We walk through a realistic end-to-end supply chain attack - a manipulated agent introduces a malicious dependency that survives code review and flows through CI/CD to downstream consumers. Attack primitives include LLM-amplified typosquatting, prompt injection via package metadata, and lateral movement into build and signing infrastructure.

Talk 2: Maurice Kraus (Research Associate / PhD Candidate at Technische Universität Darmstadt): Do We Really Need Another Forecasting Model?

Time-series foundation models promise a single solution across domains, yet often rely on architectures that may be ill-suited to latency-sensitive or specialized use cases. In contrast, xLSTM-Mixer illustrates how lighter recurrent and state-space-inspired architectures can compete with heavier transformer models in both accuracy and efficiency. Benchmarks such as QuAnTS further highlight that practical time-series modeling increasingly goes beyond forecasting, requiring capabilities such as question answering and reasoning over temporal data, often through more specialized models. Finally, recent work on finding the Zeitgeist in time-series foundation models shifts the question from prediction error alone to what these models actually learn, and how their representations transfer. This talk introduces xLSTM-Mixer and uses it to motivate a broader discussion: do we need another forecasting method, or better ways to design, understand, and evaluate models for the specific demands of each domain?

We will share more details on the talks and an agenda soon.

See you in July,
Aaqib, Marc & Fabian