In plain words
The companies building the most capable AI, often called frontier models, publish documents explaining how they decide whether a model is safe to release. Anthropic calls its version a Responsible Scaling Policy, OpenAI a Preparedness Framework, and Google DeepMind a Frontier Safety Framework. Together they are known as frontier safety frameworks.
Most follow an if-then pattern. Before release, the company tests a model for specific dangers, such as helping someone make a biological weapon or carry out a cyberattack. If the model's abilities cross a line set in advance, then stronger protections must be in place first, and in the extreme the model should not be released at all.
They began as voluntary promises in 2023. Since 2026, some laws require large developers to publish one.
Everyday example
Building codes that scale with height
A two-storey house in Pune needs basic approvals. A twenty-storey tower needs fire exits, a structural audit and a lift inspection before anyone moves in. The taller you build, the more checks apply. A frontier safety framework works the same way, except the builder writes its own code, measures its own height and decides when the next set of checks kicks in.
Where the example breaks: Building codes are written and enforced by the city. Most of a frontier safety framework is written and judged by the company itself, though new laws now require publishing one and reporting serious incidents.
How it actually works
The details differ by company, but the frameworks share four moving parts.
Pick the risks that matter most
Frameworks focus on severe, large-scale harm: chemical and biological weapons, cyberattacks, and AI that can speed up its own development or escape human control. Everyday harms like bias are handled elsewhere.
Set thresholds in advance
Each risk gets capability levels. Anthropic uses AI Safety Levels (ASL). OpenAI uses High and Critical. Google DeepMind uses Critical Capability Levels, and from April 2026 lower Tracked Capability Levels. Crossing a level triggers required safeguards.
Test before release
Teams run evaluations, red teaming and expert trials to see how close a model comes to each line, and publish summaries in model cards or risk reports.
Apply safeguards, then decide
Safeguards include classifiers that block dangerous requests, tighter security around the model's weights, and limits on access. A safety group reviews the evidence and recommends to leadership whether to deploy.
- Current versions
- As of November 2026: Anthropic RSP 3.4 (8 July 2026), Google DeepMind Frontier Safety Framework 3.1 (17 April 2026), OpenAI Preparedness Framework version 2 (15 April 2025), with a Frontier Governance Framework built on it (28 May 2026)
- First published
- Anthropic's RSP, September 2023, then OpenAI's Preparedness Framework (beta), December 2023
- Industry pledge
- Frontier AI Safety Commitments, AI Seoul Summit, 21 May 2024: signatories (now 20 companies) agreed to publish frameworks with risk thresholds
- Laws that require one
- California SB 53, in force from 1 January 2026; New York's RAISE Act from 1 January 2027. Both cover developers with over US$500 million revenue training above 10^26 operations
- EU
- The General-Purpose AI Code of Practice (10 July 2025) has a Safety and Security chapter for the most capable models under the AI Act
No Indian company appears on the Seoul commitments list as of November 2026, but Indian users rely on models these frameworks cover.
Do frontier safety frameworks actually keep AI safe?
One view
A useful start
They make companies define dangers, test for them and say in public what they will do. Laws in California and New York and the EU code have since built on them, turning voluntary habits into published duties.
The other view
Too easy to bend
Companies write, grade and revise their own rules. OpenAI's 2025 version says it may adjust requirements if a rival releases a high-risk system, and Anthropic's 2026 rewrite dropped its earlier pledge not to train or deploy models capable of catastrophic harm, and set its new roadmap as goals rather than hard commitments.
Most likely: Most observers see them as better than nothing and short of independent oversight. The open question is how much outside review and law should back them up, which is still being decided country by country.
Questions people ask
What is Anthropic's Responsible Scaling Policy?
Anthropic's frontier safety framework, first published in September 2023. It ties AI Safety Levels (ASL) to required safeguards. Version 3.0 (February 2026) added a public safety roadmap and regular risk reports; version 3.4 dates from July 2026.
What is OpenAI's Preparedness Framework?
OpenAI's framework for tracking severe risks. Version 2, from April 2025, tracks biological and chemical, cybersecurity, and AI self-improvement capabilities, with High and Critical thresholds that require safeguards before deployment or during development.
Are frontier safety frameworks legally binding?
Mostly not, they began as voluntary. But California's SB 53 (from January 2026) and New York's RAISE Act (from January 2027) require large developers to publish and follow one, with penalties, and the EU AI Act sets duties for the most capable models.
What counts as a frontier model?
Loosely, the most capable general-purpose models of the day. Laws draw a line by training compute: California and New York use more than 10^26 operations, while the EU AI Act presumes systemic risk above 10^25.
How are these frameworks different from AI regulation?
A framework is a company's own rulebook for its models. Regulation is set by governments for everyone. The two now overlap, because some laws require companies to publish their framework and report serious incidents.
Was this clear?
1 is confusing, 5 is crystal clear. One tap, no account needed.
Spotted something wrong, or know a better example? Suggest a change