Evaluating Frontier Models for Dangerous Capabilities

Research question

How should persuasion, cyber, self-proliferation and self-reasoning capabilities be evaluated?

Key findings

A dangerous-capability evaluation suite run on Gemini 1.0 models - No strong dangerous capabilities found, but early warning signs in several areas

Sources

arxiv.orghttps://arxiv.org/abs/2403.13793

Contribute

Spot an error, or know a record we should add? Members can propose additions and submit corrections.

Submit research Report a correction or request removal

Related in Research Paper Library

Evaluating Frontier Models for Dangerous Capabilities

Google DeepMind's first published dangerous capability evaluation programme, covering persuasion and deception, cyber offence, self-proliferation and self-reasoning, with the capture-the-flag and autonomy tasks that…

Dangerous capability evaluation2024

Model evaluation for extreme risks

The cross-lab paper that made dangerous-capability evaluations a pillar of frontier AI governance.

Dangerous capability evaluation2023

GPT-4 Technical Report

The GPT-4 system card set the template for frontier-model safety disclosure, including dangerous-capability red teaming.

Dangerous capability evaluation2023