Tag: LLM security
Browse our exclusive articles!
Layerwise Convergence Fingerprints for LLM Misbehavior Detection
Discover Layerwise Convergence Fingerprinting, a tuning-free method to detect runtime misbehavior in large language models with high accuracy and security.
Plug-and-Play Defense for Backdoored LLMs with TIGS
Discover TIGS, a plug-and-play defense that protects large language models from backdoor attacks without affecting performance or latency.
Layer-wise Vulnerabilities in LLMs Exposed by Mechanistic Steering
Discover how mechanistic steering reveals layer-specific vulnerabilities in LLMs, enhancing defenses against adversarial jailbreak attacks.
RouteGuard: Detecting Skill Poisoning in LLM Agents
Discover how RouteGuard detects and prevents skill poisoning in LLM agents using advanced internal-signal detection techniques.
Enhancing LLM Security with Sparse Autoencoders
Discover how integrating Sparse Autoencoders boosts Large Language Models' robustness, reducing jailbreak attacks by up to 5x without altering weights.