The Safety Report Thats Coming Soon: What Is Still Unverified

Meta's Superintelligence Lab released Muse Wonder and Muse Wonder Full on October 12, 2026 — Meta's own description calls it the company's first "frontier model," a meaningfully bigger claim than anything made about the lightweight Muse Spark model that ran the original consumer launch in September. Wonder is available now through the Meta Model API and Muse Code; Wonder Full is restricted to Muse Code subscribers. Both are described as will "soon" power the consumer Muse app, subscribers first, with no firm date given. Meta says the models are trained to avoid sharing private information without permission and to refuse harmful requests. It also says a full safety report, with additional evaluations, is coming "soon." As of this writing, five weeks after the model shipped, that report hasn't appeared.

That gap is worth sitting with rather than rushing past, because of what Wonder is actually being positioned to do.

A bigger model, a bigger claim set

Wonder isn't a modest iteration. Meta's own model card claims benchmark scores comparable to Fable 5.1 and GPT-6 Astra, and says Wonder beats all three of Fable, Opus, and Astra specifically on MemoryBench, a memory-retention benchmark. Independent numbers from Artificial Analysis put Wonder Full at 56 and Wonder at 52 on its Intelligence Index — scores that matter more here than Meta's own framing, since they come from a third party rather than the company grading its own homework. Meta is also highlighting expert-domain capability in law, finance, and health, plus stronger coding performance for building games and business apps, and two new behavioral features: Active Memory, which retains and synthesizes context across conversations, and Active Focus, which keeps the model pursuing a goal through difficulty, recovering from errors, and checking in before taking important actions.

Every one of those is a capability that raises the stakes of getting safety wrong. A model confident enough to be marketed for finance and health questions, and persistent enough to keep working on a goal without being re-prompted, is a different safety surface than a chat assistant that answers one question and stops. The original Muse launch leaned hard on its permissions architecture — the Secure VM and the Sentinel agent gating outbound actions — as the thing that made an agent this capable trustworthy. Wonder adds raw model capability on top of that architecture, and Meta has not yet published the document that would tell us whether the model's own judgment has been evaluated against the new things it's now being asked to do.

What "coming soon" actually withholds

A safety report for a model like this would typically cover things like red-teaming results against adversarial prompts in the domains Meta is specifically advertising — does Wonder's law, finance, or health advice hold up against deliberately misleading inputs, or edge cases where the "right" answer is genuinely contested? It would also typically address how Active Focus's persistence interacts with safety guardrails: a model that's designed to keep trying other approaches when a task gets hard is a model that needs a correspondingly clear description of where "trying another approach" stops and a refusal should kick in instead.

None of that exists publicly yet. What exists is a product announcement that asserts safety training occurred, without the evaluation data that would let anyone outside Meta assess whether that training actually holds up. That's a normal sequencing choice for a product launch — ship the capability announcement, follow with the evaluation paper — but it's also exactly the kind of claim that should be flagged as unverified rather than treated as settled, precisely because the promise itself is the thing standing in for the evidence right now.

This isn't just a Muse problem

It's worth saying plainly that Meta isn't alone in this gap. None of the three major personal-agent launches from this cycle has published a full safety or red-team report for their newest capability. OpenAI's usage-grace period for Dots ends in late October with no published explanation of what changes afterward. xAI's Grok Bot has been handed a materially larger role — running Grokipedia's entire content pipeline, including fact-checking and edit review, without a human editorial layer — without a published safety framework for that specific responsibility. Meta's situation is the same shape: a real capability jump, a stated intention to publish safety evaluation, and a gap between the two that's still open a month later.

The honest reading isn't that any of these companies is hiding something damning. It's that "safety report coming soon" has become the default way frontier-capability launches get sequenced in 2026 — ship first, evaluate publicly second — and that sequence puts the burden of judgment on users and reporters in the gap, not on the company.

What the original launch's safety framing already set up

It's worth remembering what Meta's safety pitch looked like before Wonder existed. The original Muse launch in September leaned on a specific, falsifiable architectural claim: nothing Muse does reaches the internet unless a separate Sentinel agent approves it, and Muse itself has no visibility into passwords or payment credentials. That's a claim about the system around the model, not the model's own judgment — and it's a claim that's held up reasonably well in the reporting gathered across this entire series so far. Wonder's safety story is a different kind of claim entirely, because it's about whether the model itself, now asked to operate in law, finance, and health, and now built to keep pursuing a goal through setbacks rather than stopping and asking, has been evaluated against the specific failure modes those capabilities invite. A permissions architecture can be airtight and a model's own judgment can still be the thing that goes wrong inside it — which is exactly why a model-level safety report matters even where the system-level architecture is already strong.

What would close this gap

The simplest fix here isn't complicated: Meta publishing the safety report it promised, with enough specificity that outside researchers can actually test the claims rather than just read them. Given that Wonder is already live in the API and already powering Muse Code subscribers' sessions, the report isn't gating a launch decision anymore — it's purely a transparency commitment at this point, which makes the delay harder to excuse the longer it continues.

Until it lands, the fair thing to say about Muse Wonder is this: Meta has made a specific, falsifiable promise about safety evaluation, the model is already live and being used for the exact domains — law, finance, health — where that evaluation would matter most, and the evidence to check the promise against isn't public yet. That's not an accusation. It's a status update, and it's the one worth tracking before any article in this series makes a confident claim about how safe Wonder actually is in practice.

Part 39 of 46 in the Meta Muse series. Previously: Law, Finance, Health: Muse Wonder's Three Expert Domains, Examined. Next: Did Muse Downloads Really Fall? Reading the TickerTrends Report Carefully. Browse more builder insights or explore AI skills at aiskill.market.