Taylan Alpay

Conductance-repair evidence graphs can boost security retrieval recall from 0.017 to 0.069, revealing the hidden potential of deterministic graph-flow methods in operational triage.

A budgeted evidence selection approach reveals that naive evaluation methods can inflate vulnerability recall metrics by over 8 times, challenging traditional assessment practices.

Layer order in cybersecurity pipelines can dramatically alter the visibility of security facts, with implications for how we design and audit layered enforcement systems.

LLM agents are surprisingly vulnerable: even small models (Qwen3-0.6B and 1.7B) struggle to maintain security boundaries against prompt injection, privacy leakage, and tool-use integrity attacks, despite various defenses.