For Heads-of · Practitioner
Frontier Safety Framework
Also known as: responsible scaling policy · frontier AI framework
A frontier AI developer's published commitment to test its models for dangerous capabilities and to apply stronger safeguards, or pause, when defined capability thresholds are crossed. Examples include Anthropic's Responsible Scaling Policy, OpenAI's Preparedness Framework and Google DeepMind's Frontier Safety Framework.
- responsible scaling
- capability thresholds
- preparedness
- frontier AI
- SB 53
How they work
Each framework names risk areas (typically chemical and biological weapons, cyber attack and a model's ability to accelerate AI development itself), defines capability thresholds in each, and says what safeguards apply once a model reaches them. Developers test before release and publish the results in system cards. The thresholds have different names in each framework, but the if-then structure is shared.
Their status
They are voluntary, and set and assessed by the developers themselves, which is their main limitation. California's SB 53 now requires large frontier developers to publish a frontier AI framework and report against it, and the EU's General-Purpose AI Code of Practice asks signatories with systemic-risk models for a comparable safety and security framework. The frameworks change often: Anthropic's policy reached version 3.4 in July 2026, Google DeepMind published version 3.1 of its framework in April 2026, and OpenAI's Preparedness Framework has been at version 2 since April 2025. Cite the version you read.
What they mean for a deployer
Read the framework and the system card for any frontier model you use: they describe what the developer tested, what it found and what it does not cover. They are evidence for third-party model and vendor due diligence, not a substitute for testing in your own context; see frontier AI.
Sources
Related Terms
Safety Case
A structured argument, supported by evidence, that a system is acceptably safe for a given use in a given context. Long established in nuclear, aviation and defence engineering, safety cases are being adopted for frontier AI as a way to make safety claims explicit and checkable.
Uplift
The additional ability an AI system gives a person to do something, compared with what they could do without it, for example using search engines and textbooks. Uplift is the measure most frontier safety thresholds use for dangerous capabilities.
General-Purpose AI Model
An AI model that displays significant generality and can competently perform a wide range of distinct tasks, however it is placed on the market. Under the EU AI Act its provider has had documentation, copyright and downstream-information duties since 2 August 2025, and a model with systemic risk carries evaluation, adversarial testing, incident reporting and cybersecurity duties on top.
AI Safety
The field of research and practice concerned with making AI systems behave as intended, avoid unintended harm and stay under meaningful human control. In 2025 the UK and US government institutes renamed themselves around security, which narrowed the official focus without removing the questions an organisation deploying AI has to answer.
Frontier AI
The most capable general-purpose AI models, at or near the leading edge. There is no single definition: the EU AI Act presumes systemic risk above 10^25 training operations, California's SB 53 and New York's RAISE Act define frontier models above 10^26, and the International AI Safety Report prefers the term general-purpose AI.
Model Card
A document describing a machine learning model's intended uses, performance, limitations and known risks, so that people deciding whether and how to use it can do so on evidence. System cards extend the idea to the deployed system and its safeguards, and EU AI Act documentation duties now require much of the same content.