**Alt text:** A cluster of white pearls inside and spilling out of an open oyster shell.

Google DeepMind’s Frontier Safety Framework (FSF) is a published policy for spotting dangerous capabilities in Gemini models before release. It defines Critical Capability Levels (CCLs), which are points at which a model could cause severe harm if no safeguards are in place. It also sets lower alert thresholds, so DeepMind gets early warning that a model is approaching a CCL. Each new frontier model is tested against both. If a threshold is crossed, DeepMind must write a response plan, tighten security on the model’s weights, add deployment safeguards and, before launch, review a safety case. The current version is 3.1, released 17 April 2026. It covers five areas: CBRN weapons, cyber attacks, harmful manipulation, machine-learning R&D, and misalignment. As of the Gemini 3.1 Pro model card, no released Gemini model has reached a CCL. Every Gemini Pro model since Gemini 2.5 Pro has crossed the cyber alert threshold.

How has the framework changed since 2024?

DeepMind has published four versions. Each one widened the scope or made the rules stricter. All of them are listed on DeepMind’s Frontier Safety page.

Version Date What changed
1.0 May 2024 The first framework set CCLs in four areas: autonomy, biosecurity, cybersecurity and ML R&D.
2.0 Feb 2025 The update assigned a recommended security level to each CCL and added a separate approach to deceptive alignment.
3.0 Sep 2025 Added a harmful-manipulation CCL and covered models that resist being shut down or modified. Safety case reviews were extended to large internal deployments of advanced ML R&D models.
3.1 Apr 2026 Added Tracked Capability Levels (TCLs), a lower tier that flags less extreme risks earlier.

Version 3.0 and the 3.1 addendum are explained in DeepMind’s post Strengthening our Frontier Safety Framework. Two of its points are easy to miss. Risk assessments must now state outright whether the remaining risk is acceptable. And safety reviews now apply to some powerful models used only inside Google, as well as to public launches.

What exactly does DeepMind test for?

  • CBRN uplift: whether the model gives meaningful help to someone trying to build chemical, biological, radiological or nuclear weapons.
  • Cyber uplift: whether the model substantially helps people carry out high-impact cyber attacks.
  • Harmful manipulation: whether the model could be misused to “systematically and substantially change beliefs and behaviors in identified high stakes contexts.” DeepMind still treats this level as exploratory.
  • ML R&D: two levels. One covers models that speed up AI research to potentially destabilising rates. The other covers models that can fully automate AI research work.
  • Misalignment: exploratory “instrumental reasoning” levels that test whether a model could work against human oversight, including attempts to avoid being shut down.

What happens when a model crosses a threshold?

Crossing an alert threshold does not stop a launch. It requires DeepMind to work out how close the model is to the CCL and to draw up a response plan. That plan has two kinds of measures:

  • Security measures make it harder for anyone to steal the model’s weights.
  • Deployment measures include safety fine-tuning, misuse classifiers and restricted access.

If a model actually reaches a CCL, a safety case review is required before external launch. The framework contains no automatic “kill switch.” Decisions go through Google’s governance process.

In practice, every published Gemini model card from 2.5 Pro onward has reported crossing the cyber alert threshold without reaching the CCL. The Gemini 3.1 Pro model card (February 2026) lists “CCL not reached” in all five areas. It says the model “has reached the alert threshold” for cyber, and that DeepMind continues “to deploy mitigations in this domain.” On manipulation, it reports a maximum odds ratio of 3.6x, the same as Gemini 3 Pro, which is below the alert threshold.

Which popular claims about the FSF are wrong?

Several widely shared explanations of the FSF describe things that are not in any version DeepMind has published:

  • There is no Green/Yellow/Orange/Red colour scale. The framework uses named capability levels within each risk area.
  • There is no published five-minute “weight zeroing” kill switch, and no weekly global shutdown drill.
  • No DeepMind interpretability tool called “Neuron Gossip” has been made public.
  • DeepMind has published no figures such as a “48% drop in severe incidents” or “11% slower shipping.”

If a source quotes any of these, it is not describing the real framework. The only reliable sources are DeepMind’s own documents and model cards.

Is the FSF still voluntary?

The FSF is Google’s own policy, but laws now require similar documents. California’s SB 53, signed on 29 September 2025, requires large frontier developers to publish a framework explaining how they assess and reduce catastrophic risk. In the EU, the General-Purpose AI Code of Practice, published 10 July 2025, has a Safety and Security chapter for providers of the most advanced models. The FSF is how Google meets both expectations. That is why regulators are interested in its structure: thresholds, early warnings and documented responses.

What does it mean for teams that use Gemini?

We use Gemini and similar models in production for scripts, shot lists, transcription and rough edits. For that kind of work, the FSF affects you in two ways.

First, the safeguards it requires show up as refusals. Requests close to cyber or weapons topics get blocked even when they’re legitimate. A thriller script with a realistic hacking scene is one example. A training video about lab safety is another. Allow time for rewording prompts, or do that part of the script by hand.

Second, the manipulation CCL is aimed at large-scale changes to beliefs in high-stakes settings such as elections or health decisions, not ordinary ad copy. Still, if a campaign touches those areas, expect stricter filtering and plan for human review before anything airs.

If you only read one primary source, read the model card for the specific model you use. It shows which thresholds were crossed and which safeguards are switched on. Read the full FSF 3.1 document on DeepMind’s Frontier Safety page when your own AI policy or a client contract needs to name a standard.

Case Studies

Clients we worked with.