Carnegie Mellon releases SafetyNudges and DecompBench
Carnegie Mellon University researchers have developed two open-source tools, SafetyNudges and DecompBench, to improve AI safety from the perspectives of both users and developers. SafetyNudges is a Chrome extension that warns users when chatbots display human-like feelings or offer medical advice, helping people remain mindful of interactions. Meanwhile, DecompBench benchmarks decomposition attacks, helping developers identify how malicious actors combine multiple small tasks across agents to bypass AI model safety protocols. Open article →