Evaluating Prompt Injection Defenses for Educational LLM Tutors: Security-Usability-Latency Trade-offs Maiorano (2026) — arXiv cs.CR/cs.AI. 📄 Full text (arXiv)
Summary
This paper addresses a fundamental tension in educational LLM deployment: tutors must be both secure (resist prompt injection attacks) and usable (not block legitimate educational interactions). The author presents a systematic evaluation methodology using a 480-query benchmark (369 injection, 111 benign) with statistically rigorous comparison.
Defense methods compared:
| Method | Bypass Rate | False Positive Rate | Latency |
|---|---|---|---|
| Proposed Multi-Layer Pipeline | 46.34% | 0.00% | 2.50ms |
| Prompt Guard (Meta) | 38.48% | 3.60% | — |
| NeMo Guardrails (NVIDIA) | 0.0% | 16.22% | 1.3s |
The proposed pipeline combines deterministic pattern filters, structural validation, contextual sandboxing, and session-level behavioral checks. Its design prioritizes pedagogical usability — zero false positives means no legitimate student queries get blocked, an essential requirement for intelligent-tutoring systems where interruptions harm learning.
NeMo Guardrails blocks all attacks but incorrectly flags ~16% of benign requests — a rate that would seriously degrade the student-experience in real tutoring sessions. Prompt Guard provides middle-ground performance.
The framework enables evidence-based guardrail selection under institutional risk and usability requirements. This directly connects to ai-tutor-safety-harms concerns and the emerging field of pedagogical-safety-rl. The latency dimension is particularly important for real-time conversational-ai-tutors-framework where response delays degrade engagement.
The paper highlights that educational settings have unique requirements: false positives are more costly than in general-purpose chatbots, because blocking a student's learning interaction carries pedagogical harm. This aligns with findings in eduframetrap-llm-sycophancy-educational-safety that educational safety requires domain-specific benchmarks.
Related Pages
- ai-tutor-safety-harms — Safety harms in AI tutoring systems
- pedagogical-safety-rl — Pedagogical safety in RL
- intelligent-tutoring — Intelligent tutoring systems
- conversational-ai-tutors-framework — Conversational AI tutoring
- eduframetrap-llm-sycophancy-educational-safety — Sycophancy as safety risk
- student-experience — Student experience with AI
- regulation — AI regulation in education