Tag: agentic AI safety
Browse our exclusive articles!
EPO-Safe: Learning AI Safety from 1-Bit Danger Signals
EPO-Safe enables AI agents to discover safety rules using minimal 1-bit danger signals, enhancing autonomous safe behavior in complex environments.
Atlas-Alignment: Scalable Interpretability for Language Models
Discover Atlas-Alignment, a framework that makes language model interpretability scalable, cost-efficient, and robust across AI systems.
Preventing AI Catastrophes: Risks of Misaligned Objectives
Explore the risks of AI misalignment and how constraining capabilities can prevent catastrophic outcomes in advanced AI systems.
Mitigating Self-Jailbreak in Large Reasoning Models Safely
Discover how Chain-of-Guardrail (CoG) improves safety in large reasoning models by preventing self-jailbreak without sacrificing reasoning power.
Estimating Tail Risks in Language Model Outputs Safely
Discover efficient methods to estimate rare harmful outputs in language models, improving safety with fewer samples and better risk prediction.