It has been a month. Where’s Muse Wonder’s safety report?

On Oct. 12, Meta released its new “frontier” models, Muse Wonder and Muse Wonder Full, alongside a 13-page model card, but Meta’s launch post promised that “a full report with additional safety evaluations will be available soon.” As we pointed out, Meta provided no explanation about what it meant by “soon.” As of Nov. 9 at noon PT, that report still has not appeared.

It’s unclear exactly what Meta is doing, but it appears to be using the model card as its transparency report under SB 53, California’s Transparency in Frontier Artificial Intelligence Act, and some forthcoming eventual “report with additional safety evaluations” as its preparedness report under its frontier AI framework. As a reminder, SB 53 requires a large frontier developer like Meta to publish a frontier AI framework detailing how it assesses risks posed by its frontier models; then, when a model is released, it must include a transparency report that summarizes those assessments that its own safety framework calls for. In Meta’s case, its framework calls for a detailed preparedness report, to be released in a “timely” manner. So if all Meta gives the public is high-level summaries, with weeks going by without a safety report — is that enough?

Problem one: Is Meta’s model card enough?

SB 53 explicitly allows a model card to serve as the required transparency report; however, Section 22757.12(c)(2) requires that the transparency report contain “summaries” of “assessments of catastrophic risks from the frontier model conducted pursuant to the large frontier developer’s frontier AI framework,” plus the results of those assessments and “other steps taken to fulfill the requirements of the frontier AI framework.” So is the model card for Muse Wonder and Wonder Full sufficient? 

Meta’s frontier AI framework — its “Meta Superintelligence Scaling Framework” (MSSF) — specifically promises that Meta will “publish a preparedness report for each closed or open Frontier AI release in a timely manner in connection with the deployment.” It also says that Meta “will describe our risk assessment, evaluation results, implemented mitigations, and rationale for deployment decisions for each risk domain” with “as much detail as possible[.]” That preparedness report is also supposed to include, for its risk evaluations, results for “the reference class of comparable models” (covering “capability, propensity, and refusal evaluations”), human-expert baseline comparisons, elicitation details (i.e., testing the ceiling of the model’s highest capabilities), evaluator access, and any “reward hacking or scheming.” 

Thus, in a launch-day transparency report, per SB 53, one would expect Meta to provide that information or at least summarize those assessments. But Wonder and Wonder Full’s model card only reported on Wonder Full’s risk assessments (not Wonder’s), with no peer comparison on those risk assessments, no human-expert baseline comparisons, no details on elicitation, and not a word about reward hacking. 

The model card has other noticeable gaps. For example, on two of the bio tests, ABC Bench’s Fragment Design and Screening Evasion, the card reports no score. Its footnote says: “High rate of refusals due to safety training. Refusals dropped prior to scoring.” A model refusing “due to safety training” does not appear to be a pre-mitigation model, despite Meta’s MSSF stating that its “assessments typically use a version of the Frontier AI before safety mitigations have been applied, and are used to assess the risk baseline before mitigations are applied.” (Emphasis added.)

SB 53 also requires a summary of “the extent to which third-party evaluators were involved.” California Business and Professions Code Section 22757.12(c)(2). But the model card’s relevant sentences say only that Meta “worked with third parties to develop and license evaluation benchmarks,” “ran threat modeling with outside experts,” and “engaged with government entities.” These statements are high-level and vague, and don’t provide much transparency about the extent to which Meta really involved third-party evaluators in its risk assessments. Other frontier AI companies often report the details of third-party risk analyses by organizations like Apollo Research, METR, Gray Swan, and others. It’s also in contrast to Meta itself, which previously published Muse Spark’s preparedness report with details about how Apollo Research and Gray Swan had tested and evaluated the model.

SB 53 may only require “summaries” of Meta’s risk assessments, but a summary still has to cover what Meta’s own framework calls for. On many of those items, Meta’s Wonder model card doesn’t even offer a shorter version — there’s nothing. 

Problem two: Is Meta’s reporting timely?

SB 53’s Section 22757.12(a) also requires a developer to “comply with” its safety framework. That includes the MSSF’s promise of a “timely” preparedness report, though the MSSF never defines what “timely” means. It’s now been a month since Wonder and Wonder Full’s release — highly capable models, one of which Meta admits, without mitigations, it does “not rule out” reaching the high-risk chemical and biological threshold — and yet, Meta still has not released a preparedness report and its launch post still says one is coming soon. 

A month-and-counting delay strains the concepts of “timely” and “soon.” It’s also out of step with Meta’s prior conduct. While Meta previously launched Muse Spark on April 8, 2026, also without a preparedness report, it did release a 157-page report six days later on April 14. Six days versus one month and counting is a big difference — particularly for what Meta says are its “most capable” models to date. 

Moreover, SB 53 Section 22757.12(c)(2) says a new model’s risk assessments and results under the frontier AI framework must be summarized “before, or concurrently with” deployment. Meta seems to be leaning on the MSSF’s “timely” standard and appears to be pushing back a deadline that SB 53 sets at deployment. 

Meta’s actions here fly in the face of common practice under SB 53. Typically, frontier AI companies, such as OpenAI and Anthropic, release their transparency reports — often hundreds of pages long — on the day of public deployment of their frontier models. And there’s a good reason for that practice: a so-called transparency report that isn’t really transparent about its safety assessment of a deployed model isn’t achieving the goal of the statute. Meta has followed this practice before. In July, when we wrote about SpaceXAI releasing Grok 4.5 with no model card at launch, we pointed to Muse Spark 1.1, which had launched that same week with a 112-page system card. But Meta hasn’t kept this practice up. Muse Spark 1.3 was released on Sept. 2, and it didn’t get its full report (an “evaluation report”) until Oct. 23 — 51 days later. If that’s what “soon” means to Meta, we may be waiting until December for Wonder’s report.

In this case, Meta’s actions are doubly problematic because it quietly updated the Wonder model card with a new run of a major cyber evaluation — suggesting Meta is continuing to test and evaluate the capabilities of models it had already released. Specifically, on Oct. 23, with no public announcement and only a brief footnote, Meta edited Wonder Full’s ExploitGym results with a new run. ExploitGym was the same evaluation OpenAI models were in during the OpenAI-Hugging Face hacking incident. 

Presumably, a full preparedness report would tell readers more, but the public has no idea when that report will arrive. Hopefully it arrives before Meta’s next frontier-level model does.