Muse Spark Safety & Preparedness Report

What the report is

This is a safety and preparedness report for Muse Spark, described as the latest large language model developed by Meta. It was written by Cristina Menghini, Peter Ney, Hamza Kwisaba and colleagues at Meta, and submitted to arXiv on 14 May 2026 as 2606.12429 ^[1]. Its purpose is to set out the evaluations that stood behind the decision to launch the model ^[1]. Muse Spark was released as the underlying model of Meta AI ^[1].

The framework it answers to

The evaluations covered the catastrophic risk domains defined in Meta's Advanced AI Scaling Framework ^[1]. That framing matters: the report is not a general audit of everything the model can do, but an assessment against a specific list of severe-risk categories that the company has committed to checking before deployment. The preparedness results span Chemical and Biological, Cybersecurity, and Loss of Control risks ^[1].

What the evaluations found

Evaluations of dual-use and high-risk capabilities identified elevated risks before mitigations were applied ^[1]. Most notably, Chemical and Biological capabilities were assessed as likely reaching the high risk category under the framework before safeguards were in place ^[1]. In other words, the model as evaluated in its unmitigated form showed enough capability in areas that could be misused for harm that it landed in, or near, the framework's highest band for those domains.

What was done about it

A multi-layered set of mitigations was implemented to address the identified risks ^[1]. The report's central claim is that, with those mitigations in place, Muse Spark shows state-of-the-art refusal across benchmarks related to hazardous workflows in chemistry and biology ^[1]. Refusal here means the model declining or redirecting requests that would involve dangerous procedures; the report treats strong refusal behaviour on such benchmarks as the evidence that the raised capability does not translate into raised risk.

On that basis the report concludes that deploying Muse Spark within Meta AI presents acceptable levels of residual risk under the framework ^[1].

Beyond the catastrophic domains

The report also discusses broader content safety and the model's behavioural profile ^[1]. These topics fall outside the catastrophic risk domains governed by the framework ^[1], but the authors include them so that the safety picture is not limited to the severest categories.

Why it matters

Reports like this are how a lab documents, before release, that it looked for the most serious kinds of risk and did something about what it found. Two features of this one are worth noting. First, the finding of elevated risk before mitigation, including Chemical and Biological capability likely reaching the high risk category, is stated plainly rather than buried; the argument for safety rests on the mitigations and on measured refusal behaviour, not on the absence of capability ^[1]. Second, the scope is explicitly bounded by the Advanced AI Scaling Framework, with content safety and behavioural questions handled separately ^[1]. For readers trying to understand how frontier models are cleared for release, the report is a concrete example of the process: define the risk domains, evaluate against them, apply layered mitigations, measure the result, and state the residual risk that remains.