CompanionSafe: From Multi-turn Safety Evaluation to Guardrails for AI Companions
cs.AI, cs.CR, cs.HC
Submitted: 2026-10-13
Updated: 2026-10-13
Comments: 40 pages, 33 figures, Code and dataset are available at https://github.com/AI45Lab/CompanionSafe
Code: https://github.com/AI45Lab/CompanionSafe
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Sources
- DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models
- MindGuard: Guardrail Classifiers for Multi-Turn Mental Health Support
- GLM-5: from Vibe Coding to Agentic Engineering
- EUDAIMONIA: Evaluating Undesirable Dynamics in AI
- GPT-4o System Card
- Llama Guard: LLM-based Input-Output Safeguard for Human-AI Conversations
- CAREBench: A Child-Safety Risk Benchmark for Language Models
- MiniMax Sparse Attention
- PseudoBench: Measuring How Agentic Auto-Research Fuels Pseudoscience
- CompanionBench: A Theory-Anchored, Real-World-Grounded Benchmark for AI Emotional Companionship
- HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusal
- EduZone: A Framework for Evaluating LLM Safety for K-12 Students and Teachers
- DARWIN: Evolving Jailbreak Adversary and Guardrail for LLM Safety Evaluation and Protection
- Qwen3 Technical Report
- AICompanionBench: Benchmarking LLMs-as-Judges for AI Companion Safety
- Long-Term Simulation Exposes Cognitive-Developmental Risks in AI Companions
- MultiTurnPSB: Evaluating Multi-Turn Jailbreak Attacks an dClassifier-Based Defenses for Medical AI Safety
- Benchmarking and Understanding Safety Risks in AI Character Platforms
- Universal and Transferable Adversarial Attacks on Aligned Language Models
Related papers
- MAVEN-T: Reinforced Heterogeneous Distillation for Real-Time Multi-Agent Trajectory Prediction
- Model Discovery Agent: LLM-assisted Bayesian experiment design for data-efficient discovery of mechanistic world models
- Agents' Last Exam
- The Clinician's Veto: Navigating Trust, Liability, and Uncertainty in Autonomous AI Prescribing
- MindHelper: Closed-Loop Embodied Mental-State Reasoning for Precision Intervention
- Incumbent Advantage: Brand Bias and Cognitive Manipulation Dynamics in LLM Recommendation Systems