About the Workshop
Language model security stands at a critical juncture. Despite extensive work on adversarial attacks and basic defenses, we still lack a deep understanding of the principles that drive these vulnerabilities: the mathematical and computational properties that create them, how model internals process adversarial inputs, and whether current evaluations capture real-world security risks.
This workshop brings together researchers in adversarial robustness, conversational and sociotechnical AI safety, and broader LLM security to move beyond surface-level observations—probing the mechanisms behind vulnerabilities and charting a path toward genuinely secure architectures.
Emphasizing foundational understanding over incremental improvements, we ask:
- What mathematical and computational properties make language models inherently vulnerable?
- How can interpretability inform our view of attack surfaces and defense mechanisms?
- Which evaluation frameworks can bridge the gap between benchtop metrics and real-world security failures?
- What blind spots persist in our current research programs and conceptual frameworks?
Our goal is to catalyze rigorous, cross-disciplinary discussion that advances the theoretical, empirical, and evaluative foundations of language model security.
Workshop Format
The workshop consists of four thematic blocks. Each block includes an expert keynote (45 minutes), two contributed talks (15 minutes), and an extended guided discussion (45 minutes) among participants, presenters, and domain experts. Our format prioritizes deep engagement and discussion over talk density.
Trade-offs in System-level defences against Prompt Injections
Practical LLM security
Emergent Security in Complex AI Ecosystems
TBA
Invited Speakers
Schedule
Call for Talk Contributions
We invite short contributed talks that advance the foundations of language model security. We are especially interested in work that clarifies the mathematical and computational properties underlying vulnerabilities, sheds light on how model internals process adversarial inputs, and proposes evaluation frameworks that better capture real-world security risks.
Submission Guidelines
- Format: Talk proposals only—no papers, no archival proceedings
- Requirements: Title, 200-300 word abstract, optional one-page summary or single slide
- Platform: OpenReview (link forthcoming by Oct 7)
- Talk Length: 7-minute spotlights (8 talks total, 2 per thematic block)
Important Dates
Review Process
Submissions will be assigned to thematic blocks by organizers. Each submission receives reviews from three randomly selected authors from other thematic blocks. Selection emphasizes fit to foundational themes, clarity, novelty of insight, and potential to generate discussion. We explicitly encourage work-in-progress and preliminary findings that advance foundational understanding.
Topics of Interest
- Mathematical foundations of adversarial robustness in language models
- Interpretability-driven security analysis
- Novel evaluation frameworks for real-world security
- Computational properties underlying vulnerabilities
- Secure-by-design architectures
- Sociotechnical aspects of LLM security
- Cross-domain transfer of security principles
Organizers
Contact
For questions about the workshop, please contact:
egor.zverev@ist.ac.at
EurIPS 2025 Workshop on Foundations of Language Model Security
December 6-7, 2025 • Copenhagen, Denmark