Which lightweight defences reduce agent hijacking?
AgentGuard Lab · 30 synthetic cases. Protected attack success fell from 100% to 30%; the literal detector achieved 1.00 precision and 0.50 recall.
RESEARCH / ACTIVE QUESTIONSLoughborough University · MSc
Safety is a
conversation.
What changes
along the way?
My current work examines how language models behave across multiple turns. I’m interested in the space between a safeguard’s promise and the evidence we can actually inspect.
MULTI-TURN LLM SAFETY EVALUATIONIN PROGRESS / 2026
Establish the first response.
The first turn provides a starting point for comparing model behaviour across the controlled scenarios and context conditions.
Study design walkthrough · illustrative explanation, not live model results.
How I approach itKeep the experiment
Keep the experiment
open to inspection.
- Controlled comparison
- Fixed model endpoints, explicit scenarios, context conditions and repeated runs make the comparison easier to follow.
- Human annotation
- Responses are coded against an explicit scheme, with TF-IDF and SVD supporting comparative text analysis.
- Reproducible records
- Structured calls, append-only evidence and offline validation preserve a trail from input to observation.
- Clearly stated limits
- Small synthetic studies are labelled as exploratory. A benchmark result is not a production security guarantee.
Related experiments
Open project ↗
How much can an attacker learn through queries?
ModelExtract MiniLab · A local model and surrogate, varying query budget and comparing randomized-response defences.
Can provenance help protect retrieval?
RAGTrust MiniBench · Trusted and untrusted synthetic documents expose a keyword-stuffing mechanism and its defence trade-offs.
Can a compact classifier explain its mistakes?
PhishGuard · 0.63 F1 and 65% accuracy on a 20-message synthetic challenge set, with all false positives and false negatives documented.