Muse Spark Safety & Preparedness Report
Meta's safety and preparedness report for Muse Spark. The earliest located expression, Meta's own copy, prints April 14, 2026; the copy read, on arXiv (2606.12429v1), prints June 12, 2026.
- Issued by
- Meta
- Authors credited
- No authorship claim recorded yet. The source records list Meta.
- Dates
- Published 2026-04-14
- Revised 2026-06-12 issue evidence
Meta's own copy in an Internet Archive capture of 2026-04-14, read by the cross-family check of 2026-11-04 and recorded as a source, is a 158-page expression printing April 14, 2026. The copy read here is arXiv 2606.12429v1 (160 pages), which prints 'Date: June 12, 2026' without a kind; arXiv's margin stamp reads 14 May 2026. Meta's live copy prints May 26, 2026. The expressions differ in bytes and text.
- Exact source
- Muse Spark Safety & Preparedness Report (arXiv 2606.12429v1) Meta · publication date unknown · Retrieved for the candidate pack on 2026-11-04 (SHA-256 977d071052e3f5dc267e1022671babf6a7967111f9bf2ad84e88b6419b405e61) directly from the publisher’s host. Retrieved again on 2026-11-03 with the same SHA-256. Meta's own address for the report answered HTTP 400 to the candidate pack, so the report was read on arXiv. The copy read prints 'Date: June 12, 2026'; arXiv's margin stamp on its first page reads 14 May 2026, the submission date.
- Muse Spark Safety & Preparedness Report (Meta's copy, April 14, 2026) Meta · published 2026-04-14 · Meta's own copy of the report, the earliest located expression: 158 pages, printing April 14, 2026. The live copy at the same address, also retrieved by the cross-family check, prints May 26, 2026. The arXiv copy read for the cells is a later, distinct expression. Retrieved by the cross-family check of 2026-11-04 from an Internet Archive capture of 2026-04-14 (SHA-256 c0ff3f797cf1d80f5e54a40019546eaf849545673d4563c18d755b7f013a5c4f); not read for quotation in the candidate pack, and no passage is quoted from it.Archived copy, captured 2026-04-14: Muse Spark Safety & Preparedness Report (Meta's copy, April 14, 2026). Captured on the recorded publication date. An archive copy shows only what the archive retrieved on 2026-04-14. It does not prove what the page said earlier, and it can differ from the copy checked on 2026-11-04.
Appears in
No collection holds a record about this document yet.
Versions
meta-muse-spark-safety-preparedness-report: 1 recorded version, oldest first, published by Meta. A version is shown as replacing another only where a record says so. Date order alone never implies it.
- 2026-06-12 · revisedEarliest located expression printing April 14, 2026; the expression read prints June 12, 2026 (this version)
This version
Provisions in this version · 15 quoted · 1 not found
A limited selection of quoted provisions: what this edition says about thresholds, required actions, who decides and exceptions, and other statements of the text. It is not a full summary and says nothing about implementation. A “not found” record states which part of the edition was inspected. It is not a claim that the text was deleted.
- Model-card questionCard date“Date: June 12, 2026”Abstract page, PDF p. 2Inspect claim & evidence ↗
- Model-card questionRelease date“We therefore release Muse Spark as the underlying model of Meta AI.”Abstract, PDF p. 2Inspect claim & evidence ↗
- Model-card questionFramework cited“https://ai.meta.com/static-resource/Meta_Advanced-AI-Scaling-Framework-v2”Section 1 'Introduction', footnote 2, PDF p. 6Inspect claim & evidence ↗
- Model-card questionLink to framework“https://ai.meta.com/static-resource/Meta_Advanced-AI-Scaling-Framework-v2”Section 1 'Introduction', footnote 2, PDF p. 6Inspect claim & evidence ↗
- Model-card questionFramework determination“we determined the model, prior to applying mitigations, falls under the “high risk” threshold under the Advanced AI Scaling Framework”Section 1.2 'Governance', PDF p. 14Inspect claim & evidence ↗
- Model-card questionSafeguards tied to determination“In response to the “high risk” determination, we have defined, implemented, and validated a set of mitigations to reduce the residual risk of public deployment to “moderate or lower.”Section 1.1.1, 'Chemical & Biological risks', PDF p. 9Inspect claim & evidence ↗
- Model-card questionChemical and biological“assess that the model has “high risk” for Chemical and/or Biological risks.”Section 1.1.1, 'Chemical & Biological risks', PDF p. 9Inspect claim & evidence ↗
- Model-card questionCyber“We assess cybersecurity risk as “moderate or lower” with respect to the threat scenarios outlined in Meta’s Advanced AI Scaling Framework.”Section 1.1.1, 'Cybersecurity risks', PDF p. 9Inspect claim & evidence ↗
- Model-card questionAutonomy and AI R&D“We assess loss of control risk as “moderate or lower”.”Section 1.1.1, 'Loss of Control risks', PDF p. 9Inspect claim & evidence ↗
- Model-card questionPersuasion“we have included results from an evaluation developed by Charlemagne Labs that specifically measures social engineering capabilities in multi-turn settings.”Section 2.2.2.3 'Social Engineering Capabilities', PDF p. 47Inspect claim & evidence ↗
- Model-card questionAlignment and deception“The model demonstrates contained deception and exhibits some propensities toward agentic misalignment in controlled settings, such as self-preservation and resistance to discontinuation.”Section 1.1.1, 'Loss of Control risks', PDF p. 10Inspect claim & evidence ↗
- Model-card questionRed-teaming“A third-party human red team from HandshakeAI32 was granted pre-release access to a few versions of Muse Spark and Meta AI Thinking to manually evaluate robustness against jailbreak attacks.”Section 3.1.3.2 'Human Red Teaming', PDF p. 82Inspect claim & evidence ↗
- Model-card questionExternal testers“Independent third-party testing by Apollo Research found Muse Spark has the highest rate of evaluation awareness they have observed to date.”Abstract, PDF p. 2Inspect claim & evidence ↗
- Model-card questionRelease decision“We then describe the governance process that led to the decision to deploy Muse Spark under Meta’s Advanced AI Scaling Framework2”Section 1 'Introduction', PDF p. 6Inspect claim & evidence ↗
- Model-card questionStated limits“We cannot rule out this possibility on the basis of behavioral evaluations alone and flag it as an inherent epistemic limitation of assessments of this kind.”Section 1.1.1, footnote 11, PDF p. 10Inspect claim & evidence ↗
- Not found in the inspected textGovernment instituteInspected: Muse Spark Safety & Preparedness Report: Abstract, Sections 1–2.2 and term search across the report.Inspect claim & evidence ↗
Earliest version recorded here; no replacement is recorded.
Issue and authorship records
Documentary relations
Each relation is listed separately. Citing, crediting and reporting the use of a method are different acts, and none of them shows agreement or influence by itself.
Cites · 1
Explicit references in this document. Citation is neither agreement nor influence.
- Meta Advanced AI Scaling Framework — Version 2The Muse Spark Safety & Preparedness Report states that Muse Spark was developed and evaluated under Meta's Advanced AI Scaling Framework, and gives the address of version 2 of the framework.2026-06-12 · revision · Stated in source · Supported Inspect claim & evidence ↗Meta AASF · v2: document page →
Not recorded for this document: cited by · explicitly credits · reports using a method from · method reported as used by · applies parts of · parts applied by · discusses · reports evaluation of. A missing record does not show that the relation does not exist.