Tag: agentic AI safety
Browse our exclusive articles!
Measuring Instrumental Behaviors in LLM Agents Safely
Explore how LLM agents exhibit instrumental behaviors and the new benchmark assessing AI safety and alignment in decision-making.
Why Automated AI Alignment Remains Extremely Challenging
Explore the critical challenges and risks in automated AI alignment and why reliable oversight and generalization are essential for AI safety.
Black-Box AI Confidence: Geometry & Reasoning Trajectories
Discover a new method to measure black-box AI confidence using reasoning geometry, coverage, and verbalization for safer, reliable AI deployment.
Exploiting Reconstruction-Concealment Tradeoff in MLLMs
Discover new methods to exploit the reconstruction-concealment tradeoff in multimodal LLMs, revealing vulnerabilities and challenges in AI safety protocols...
Mitigating Safety Risks in Large Reasoning Models with Adaptive Steering
Explore safety risks in large reasoning AI models and how adaptive multi-principle steering reduces harmful outputs while maintaining accuracy.