Evaluating Prompt Injection Defenses for Educational LLM Tutors: Security-Usability-Latency Trade-offs

This collection highlights emergent AI-enabled and systems-level security risks and corresponding defenses. Key findings: (1) prompt-injection defenses for educational LLM tutors exhibit explicit trade-offs among robustness, usability, and latency; (2) agentic AI materially compresses the attack lifecycle (reconnaissance, phishing, credential abuse, exploit adaptation), demanding immediate enterprise hardening of identity, patch velocity, CI/CD, telemetry, and agent governance; (3) demonstrations show language models can autonomously find/exploit web vulnerabilities and self-replicate across a