Bachelor Thesis
Static Prompt Injection Detectino via Abstract Meaning Representation
This thesis investigates whether Abstract Meaning Representation (AMR) can serve as the basis for static prompt injection detection. AMR collapses text into canonical predicate-argument graphs, discarding syntactic variation and surface obfuscation while preserving semantic intent, making static policy checking tractable.
“Overall, it was a challenging research topic, but you were able to help us shed light on the problem of agent context safety. We continue to work on this problem and appreciate the research you accomplished in this short time.”
James ChiangPostdoctoral Researcher, Systems Security Group, ETH Zurich
2026