Medium

DeepMind Safety Research

DeepMind Safety Research

·

May 29

Testing Gemini models for scheming tendencies

By Victoria Krakovna, David Lindner, Sebastian Farquhar and Rohin Shah

Testing Gemini models for scheming tendencies

DeepMind Safety Research

·

Apr 1

Predicting When RL Training Breaks Chain-of-Thought Monitorability

Why RL training teaches models to hide their reasoning, and a conceptual framework to predict when it happens.

Predicting When RL Training Breaks Chain-of-Thought Monitorability

DeepMind Safety Research

·

Nov 3, 2025

Consistency Training Could Help Limit Sycophancy and Jailbreaks

Authors: Alex Irpan* and Alex Turner*, Mark Kurzeja, David Elson, and Rohin Shah

Consistency Training Could Help Limit Sycophancy and Jailbreaks

DeepMind Safety Research

·

Jul 8, 2025

Evaluating and monitoring for AI scheming

By Victoria Krakovna, Scott Emmons, Erik Jenner, Mary Phuong, Lewis Ho, and Rohin Shah

Evaluating and monitoring for AI scheming

DeepMind Safety Research

·

Apr 8, 2025

An Approach to Technical AGI Safety and Security

We have written a paper on our approach to technical AGI safety and security. This post is a copy of the extended abstract, which…

An Approach to Technical AGI Safety and Security

DeepMind Safety Research

·

Mar 26, 2025

Negative Results for Sparse Autoencoders On Downstream Tasks and Deprioritising SAE Research…

Lewis Smith, Sen Rajamanoharan, Arthur Conmy, Callum McDougall, Janos Kramar, Tom Lieberum, Rohin Shah, Neel Nanda

DeepMind Safety Research

·

Feb 14, 2025

Introducing our short course on AGI safety

We are excited to release a short course on AGI safety for students, researchers and professionals interested in this topic. The course…

Introducing our short course on AGI safety

DeepMind Safety Research

·

Jan 31, 2025

Steering Gemini using BIDPO vectors

By Alex Turner and Mark Kurzeja

Steering Gemini using BIDPO vectors

DeepMind Safety Research

·

Jan 23, 2025

MONA: A method for addressing multi-step reward hacking

MONA enhances safety when we train an AI system to perform some task that takes multiple steps. Training an AI with MONA reduces its …

MONA: A method for addressing multi-step reward hacking

DeepMind Safety Research

·

Dec 23, 2024

Human-AI Complementarity: A Goal for Amplified Oversight

How do we ensure humans can continue to oversee increasingly powerful AI systems? We argue that achieving human-AI complementarity is key.

Human-AI Complementarity: A Goal for Amplified Oversight

DeepMind Safety Research

DeepMind Safety Research

We research and build safe AI systems that learn how to solve problems and advance scientific discovery for all. Explore our work: deepmind.google

Help

Status

About

Careers

Press

Blog

Store

Privacy

Rules

Terms

Text to speech