Developing Gemini 4 Argon responsibly and safely
DeepMind describes how it is safely deploying Gemini 4 Argon, a frontier cyber-capable model, addressing misuse, misalignment, and prompt injection through staged limited access and guardrails.
Research / Sources
Google DeepMind Safety Research
Safety research from Google DeepMind. deepmindsafetyresearch.medium.com
Responsibility & Safety (19)ai (1)ai-assistance (1)ai-training (1)alignment-problem (1)amplified-oversight (1)complementarity (1)google-gemini (1)hybridization (1)jailbreak (1)language-model (1)sycophancy (1)
Focus All
Topic All
Date Any time
Piloting the world's first double-blind AI evaluations
Google DeepMind pilots double-blind external evaluation of a Gemini Flash Lite model on confidential benchmarks in a cryptographic environment, so the model cannot train on test questions.
Our approach to bioresilience
Google DeepMind and Isomorphic Labs outline a bioresilience program that aims to stop threat actors misusing their models for biology while using AI for outbreak detection and response.
Securing the future of AI agents
Google DeepMind presents its AI Control Roadmap, a defense-in-depth system that treats internal agents as potentially misaligned and grants permissions based on verified behaviour.
Investing in multi-agent AI safety research
Announces up to $10M in funding for research on safety risks from emergent collective behaviour in large-scale multi-agent AI systems, which current single-model evaluations do not cover.
Testing Gemini models for scheming tendencies
Google DeepMind introduces Gram, an automated auditing framework, and honeypot evaluations to test whether Gemini coding agents would sabotage their safeguards.
Predicting When RL Training Breaks Chain-of-Thought Monitorability
Introduces a conceptual framework predicting when RL training pressure makes chain-of-thought unmonitorable, with empirical tests on a coin-flip task and obfuscated reward hacking.}
Protecting people from harmful manipulation
DeepMind releases an empirically validated toolkit and human-participant methodology for measuring AI's capacity for harmful manipulation in areas like finance and health.
Gemma Scope 2: helping the AI safety community deepen understanding of complex language model behavior
Releases Gemma Scope 2, an open suite of interpretability tools (SAEs and related) for all Gemma 3 models, aimed at safety work on jailbreaks, sycophancy and auditing agents.
Deepening our partnership with the UK AI Security Institute
Announces a Google DeepMind and UK AISI research partnership covering frontier model evaluations, chain-of-thought monitorability, and social and economic impact research.
Consistency Training Could Help Limit Sycophancy and Jailbreaks
Proposes consistency training, where models learn to ignore user-bias and jailbreak cues by matching their own cue-free responses, including activation-level ACT; reports improved robustness on Gemini 2.5 Flash.
Strengthening our Frontier Safety Framework
Google DeepMind's third Frontier Safety Framework adds a harmful-manipulation critical capability level and addresses misaligned models that might resist operators' control or shutdown.
Evaluating and monitoring for AI scheming
DeepMind reviews evaluations of frontier models' stealth and situational awareness as scheming preconditions, and stress-tests chain-of-thought monitoring as a defense against future scheming.
Advancing Gemini's security safeguards
DeepMind white paper on defending Gemini 2.5 against indirect prompt injection in agentic tools, using automated red-teaming to probe and harden defenses.
An Approach to Technical AGI Safety and Security
Google DeepMind outlines its technical approach to preventing severe harms from AGI, covering misuse and misalignment risks in foundation models and prioritising mitigations for foreseeable capabilities.
Evaluating potential cybersecurity threats of advanced AI
Google DeepMind presents a framework for evaluating frontier AI's offensive cyber capabilities across the attack chain, grounded in real-world data, to help prioritise defenses under its Frontier Safety Framework.
Taking a responsible path to AGI
DeepMind outlines its approach to safely developing AGI, emphasising technical safety, proactive risk assessment, readiness planning, and collaboration with the wider AI community.
Negative Results for Sparse Autoencoders On Downstream Tasks and Deprioritising SAE Research…
GDM interpretability team reports that sparse autoencoders underperformed linear probes at detecting harmful intent out-of-distribution, so they are deprioritising SAE research.
Introducing our short course on AGI safety
A short introductory course on AGI alignment covering misalignment risk arguments, specification gaming and goal misgeneralization, DeepMind's amplified oversight approach, and institutional safety practices.
Updating the Frontier Safety Framework
DeepMind updates its Frontier Safety Framework with security levels for critical capabilities, consistent deployment mitigations, and an approach to deceptive alignment risk.
Steering Gemini using BIDPO vectors
DeepMind reports that BIDPO-trained steering vectors boosted TruthfulQA in Gemini 1.5 but failed to beat stronger prompt-based baselines after a model update, so the project ended.
MONA: A method for addressing multi-step reward hacking
Introduces MONA, a training method that uses myopic optimization with non-myopic human approval to reduce multi-step reward hacking and hidden agendas in post-trained LLM agents.
Human-AI Complementarity: A Goal for Amplified Oversight
Argues that human-AI complementarity, via rater assistance and hybridization protocols, can produce stronger oversight signals than humans or AI raters alone, addressing scalable oversight of advanced AI.
Mapping the misuse of generative AI
Analyzes nearly 200 media reports of real-world generative AI misuse (2023–2024), categorizing exploitation tactics and compromise patterns to inform safety evaluations and mitigations.
Introducing the Frontier Safety Framework
Google DeepMind introduces its Frontier Safety Framework, a protocol for identifying severe model-level risks such as advanced agency and cyber capabilities, and mitigating them via capability evaluations.
The ethics of advanced AI assistants
DeepMind's ethics foresight paper surveys ethical and societal risks of advanced AI assistants, covering value alignment, safety and misuse, economic and information impacts, and access.
Evaluating social and ethical risks from generative AI
Proposes a three-layer context-based framework (capability, human interaction, systemic impact) for evaluating social and ethical risks of generative AI, illustrated with a misinformation case study.
Exploring institutions for global AI governance
White paper by DeepMind and collaborators examining which international institutions could provide governance functions for managing global risks and benefits of frontier AI.
An early warning system for novel AI risks
Proposes a framework for evaluating general-purpose models for dangerous capabilities such as manipulation, deception and cyber-offense, and for alignment, to flag extreme risks early.