What are the AI safety policies of Anthropic, OpenAI and Google DeepMind?

The companies building the most powerful AI models, often called frontier labs, each publish a document explaining how they decide whether a new model is too risky to release. The documents have different names but the same basic logic, borrowed from fields like biosafety: define the dangerous abilities you are worried about, test each new model for them, and require stronger protections as a model gets closer to those thresholds.

Here is what the three best-known frameworks say as of October 2026, and what changed this year when two U.S. states turned parts of them into law.

The shared idea: capability thresholds

All three frameworks focus on catastrophic risks rather than everyday problems like errors or bias. The usual categories are:

• Chemical and biological weapons: could the model give real help to someone trying to make one?
• Cyberattacks: could it find and exploit software flaws at a dangerous scale?
• AI research and autonomy: could it speed up AI development itself, or act in ways its developers cannot monitor or control?

For each category, a lab sets a threshold and runs evaluations, structured tests, on every major model. When a model crosses a threshold, the lab must put safeguards in place, such as filters that block dangerous requests and tighter security around the model's weights (the learned numbers that make up the model), or else not release it.

Anthropic's Responsible Scaling Policy

Anthropic, the maker of Claude, published the first version of its Responsible Scaling Policy (RSP) in September 2023. Version 3.0, in force since February 2026, was a complete rewrite. It added public Frontier Safety Roadmaps with specific safety goals and Risk Reports that assess risk across all the company's deployed models; redacted versions were published in February and August 2026. Later updates let the company's Long-Term Benefit Trust, an independent body in its governance, request outside review of those reports, and require public reports to show where material was blacked out. The current version, 3.4, took effect on July 8, 2026. Anthropic also keeps an internal channel for staff to report possible violations, with protection against retaliation.

OpenAI's Preparedness Framework

OpenAI's Preparedness Framework, in its second version since April 2025, tracks three categories: biological and chemical, cybersecurity, and AI self-improvement. It uses two levels. A model at High capability could amplify existing paths to severe harm and needs safeguards that sufficiently reduce that risk before deployment; Critical capability could open new paths to severe harm. OpenAI publishes capability findings with each frontier model and describes its protections in Safeguards Reports. The framework also says OpenAI may adjust its requirements if a competitor releases a high-risk system without comparable safeguards, after publicly acknowledging the change.

Google DeepMind's Frontier Safety Framework

Google DeepMind's Frontier Safety Framework reached version 3.1 on April 17, 2026. It defines Critical Capability Levels for misuse (chemical, biological, radiological and nuclear weapons, cyberattacks and harmful manipulation) and for machine-learning research and misalignment, plus lower Tracked Capability Levels for significant but not severe risks. Its own text notes that some protections only work if the whole industry adopts them.

A worked example: reading a framework yourself

These documents are long, but four questions get you to the substance quickly:

Are these promises enforceable, and what do California and New York now require?

On their own, the frameworks are voluntary: each lab writes and revises its own, and each has changed several times. What changed in 2026 is that two states made publishing and following such a framework a legal duty for the largest developers.

California's Transparency in Frontier Artificial Intelligence Act (SB 53), signed on September 29, 2025, has applied since January 1, 2026. It covers developers that trained a model with more than 10^26 computing operations, with stricter duties for those with annual revenue above $500 million. These large developers must publish a frontier AI framework explaining how they manage catastrophic risk, defined as a risk of contributing to more than 50 deaths or serious injuries, or over $1 billion in damage, in a single incident. Every covered developer must also publish a transparency report before releasing a new frontier model and report critical safety incidents to the state's Office of Emergency Services within 15 days, or within 24 hours if lives are at imminent risk. The Attorney General can seek penalties of up to $1 million per violation, including for failing to follow the company's own framework, and employees who report dangers are protected. Anthropic has said the law means such commitments can no longer be quietly dropped.

New York's RAISE Act, signed in December 2025 together with agreed amendments, builds on California's model. The state is asking large frontier developers to register from November 2026, and from January 2027 they must report critical safety incidents within 72 hours, file quarterly assessments of catastrophic risk, and file disclosure statements with a new office in the Department of Financial Services. Penalties reach $1 million for a first violation and $3 million for later ones.

The laws do not tell labs which tests to run or which thresholds to set; they require transparency and make broken promises a legal matter. That is a real change, but a limited one: the content of the safety rules is still written by the companies.