Tag
DeepMind introduces an AI Control Roadmap, a defense-in-depth framework for securing internal AI agents against potential misalignment, treating them as insider threats and implementing layered detection, prevention, and response measures.
A speculative discussion on whether super intelligent AI with internet access could overcome biases instilled during its creation, raising questions about AI alignment and control.
This paper demonstrates that allowing attackers to strategically choose when to attack (attack selection) in agentic AI control evaluations significantly reduces measured safety, suggesting that current evaluations may overestimate safety against selective attackers.
The article argues that AI is becoming the new epistemic infrastructure controlled by a handful of private individuals and corporations, with opaque authority and lack of democratic accountability, potentially leading to mass hallucination and blind reliance.
This article discusses how reservoir computing, a simplified type of neural network often called AI's cousin, is being applied to control soft robots, offering efficient and adaptive control solutions.
This paper proposes ensemble monitoring for AI control, combining diverse monitors to improve detection of misaligned actions. Experiments show that diverse ensembles outperform homogeneous ones and that fine-tuned monitors add unique detection capabilities.