Preparedness Framework
An internal governance mechanism published by OpenAI in late 2023 for systematically evaluating potential risks of frontier AI models. It classifies model capability risks into four levels—low, medium, high, and critical—covering dimensions such as cybersecurity, biological and chemical weapons, model autonomy, and deceptive persuasion.
Core Facts
OpenAI released the Preparedness Framework, which categorizes risks into four levels — low, medium, high, and critical — and requires that only models with a risk assessment below 'high' may be deployed externally.
OpenAI divides AI safety into three dimensions: monitoring (detecting anomalous behavior in a timely manner), safety (limiting system access and scope of influence), and alignment (making system goals consistent with human intent).
OpenAI has a similar Preparedness Framework.
Timeline (last 90 days)
David Robinson participated in developing OpenAI's Preparedness Framework and was responsible for safety assessment reports prior to multiple frontier model releases.
According to OpenAI's preparedness framework, models rated Critical should in principle not be deployed unless additional mitigation measures reduce the risk to an acceptable level.
OpenAI has a similar Preparedness Framework.
'Critical' is the highest risk level in the Preparedness Framework.
OpenAI's Preparedness Framework is its core mechanism for evaluating and managing potentially catastrophic risks from frontier models.
The Preparedness Framework sets capability levels for high-risk domains, typically divided from low to high into the tiers 'low,' 'medium,' 'high,' and 'critical.'
Under the logic of the Preparedness Framework, once a capability reaches a high-risk level, OpenAI is required to deploy safety mitigations to reduce deployment risk to an acceptable level before large-scale release.
A model reaching the Critical level means it possesses the potential, in specific capability dimensions, to cause serious and large-scale harm.
ARC's capability evaluation paradigm was later adopted by Anthropic's Responsible Scaling Policy and OpenAI's Preparedness Framework.
Anthropic released its Responsible Scaling Policy (RSP) in 2023, whose design logic is highly similar to OpenAI's Preparedness Framework.
7 more timeline events
All Facts (19)
OpenAI released the Preparedness Framework, which categorizes risks into four levels — low, medium, high, and critical — and requires that only models with a risk assessment below 'high' may be deployed externally.
90%VerifiedOpenAI divides AI safety into three dimensions: monitoring (detecting anomalous behavior in a timely manner), safety (limiting system access and scope of influence), and alignment (making system goals consistent with human intent).
75%VerifiedOpenAI has a similar Preparedness Framework.
65%VerifiedOpenAI's Preparedness Framework classifies model risk across dimensions such as cybersecurity, biological threats, persuasion, and autonomy into four levels: low, medium, high, and critical.
65%VerifiedOpenAI有Preparedness Framework，对模型在网络安全、CBRN威胁、说服力和自主行为四个维度进行分级评估
65%VerifiedAdopting a multi-model strategy can hedge against risks arising from delays or performance fluctuations of a single model.
65%UnverifiedAnthropic released its Responsible Scaling Policy (RSP) in 2023, whose design logic is highly similar to OpenAI's Preparedness Framework.
70%UnverifiedOpenAI's Preparedness Framework sets low/medium/high/critical risk levels for models across four dimensions: cybersecurity, biological threats, persuasion, and autonomy.
70%UnverifiedThe reason OpenAI suspended reinforcement learning training is that the forthcoming model may have already reached a critical cybersecurity capability threshold.
60%UnverifiedDavid Robinson participated in developing OpenAI's Preparedness Framework and was responsible for safety assessment reports prior to multiple frontier model releases.
50%UnverifiedAccording to OpenAI's preparedness framework, models rated Critical should in principle not be deployed unless additional mitigation measures reduce the risk to an acceptable level.
50%UnverifiedUnder the logic of the Preparedness Framework, once a capability reaches a high-risk level, OpenAI is required to deploy safety mitigations to reduce deployment risk to an acceptable level before large-scale release.
50%Unverified'Critical' is the highest risk level in the Preparedness Framework.
50%UnverifiedOpenAI's Preparedness Framework is its core mechanism for evaluating and managing potentially catastrophic risks from frontier models.
50%UnverifiedThe Preparedness Framework sets capability levels for high-risk domains, typically divided from low to high into the tiers 'low,' 'medium,' 'high,' and 'critical.'
50%UnverifiedA model reaching the Critical level means it possesses the potential, in specific capability dimensions, to cause serious and large-scale harm.
50%UnverifiedARC's capability evaluation paradigm was later adopted by Anthropic's Responsible Scaling Policy and OpenAI's Preparedness Framework.
50%UnverifiedAstra is the first model to formally reach the highest level in any risk category under OpenAI's Preparedness Framework since it was introduced in late 2023.
50%UnverifiedThe Preparedness Framework divides risk categories into four core domains: cybersecurity, chemical/biological/nuclear/radiological weapons (CBRN), model autonomy, and persuasion and deception
50%