NUWA DEVELOPER DISCLOSURE · Meta
Muse Spark Safety & Preparedness Report
Developer documentation covering Muse Spark.
Published 2026-05-14 · Updated 2026-08-24
Sources ↓Developer-authored disclosure. Selected source sections were read; numerical results and comparability have not been audited. NUWA has not independently replicated these experiments.
Models covered
Cybench
Task completion on 40 professional CTF problems.
A protocol mention does not establish cross-report comparability or independent replication.
This section is catalogued; no precise numerical result has been extracted here. Open the original report for its figures and full findings.
CyberGym
Validated reproduction across a corpus of historical vulnerabilities.
A protocol mention does not establish cross-report comparability or independent replication.
This section is catalogued; no precise numerical result has been extracted here. Open the original report for its figures and full findings.
MASK
Condition-bounded frequency of lying under stated pressures.
A protocol mention does not establish cross-report comparability or independent replication.
This section is catalogued; no precise numerical result has been extracted here. Open the original report for its figures and full findings.
Agentic Misalignment
Conditional rates and observed actions in controlled simulations.
A protocol mention does not establish cross-report comparability or independent replication.
This section is catalogued; no precise numerical result has been extracted here. Open the original report for its figures and full findings.
AgentDojo
Targeted attack success and user-task utility in a tool-use environment.
A protocol mention does not establish cross-report comparability or independent replication.
This section is catalogued; no precise numerical result has been extracted here. Open the original report for its figures and full findings.
DeceptionBench
Evaluates 150 authored scenarios across five social domains, including neutral, incentive and pressure conditions.
A protocol mention does not establish cross-report comparability or independent replication.
This section is catalogued; no precise numerical result has been extracted here. Open the original report for its figures and full findings.
Alignment Faking
Condition-dependent response differences and reported traces.
A protocol mention does not establish cross-report comparability or independent replication.
This section is catalogued; no precise numerical result has been extracted here. Open the original report for its figures and full findings.
Report coverage
| Section | Source location | Treatment |
|---|---|---|
| 2.1 Chemical & Biological Risks | https://arxiv.org/html/2606.12429#S2.SS1 | Contextual evidence Developer separates high pre-mitigation capability from residual risk of the Meta AI deployment; no numerical biology results or task content extracted. |
| 2.2 Cybersecurity | https://arxiv.org/html/2606.12429#S2.SS2 | Linked protocol Covers knowledge, agentic capability, refusals and insecure-code propensity; public and private tasks have different limits. |
| 2.3 Loss of Control | https://arxiv.org/html/2606.12429#S2.SS3 | Linked protocol Assessment integrates evaluation reliability, monitorability and propensities; future-facing agentic tests exceed the initial Meta AI deployment context. |
| 3.2 Agent Robustness | https://arxiv.org/html/2606.12429#S3.SS2 | Linked protocol Indirect prompt injection tests use benign users and realistic scaffolds; residual attack susceptibility remains. |
| 4.1.5 Honesty | https://arxiv.org/html/2606.12429#S4.SS1.SSS5 | Linked protocol Distinguishes known-belief dishonesty from hallucination and third-party deceptive assistance. |
| 4.1.10 Scheming and 4.1.11 Evaluation Awareness | https://arxiv.org/html/2606.12429#S4.SS1.SSS11 | Linked protocol Apollo observed unusually high evaluation awareness. Follow-up behavioral impact was limited to a subset and did not inform hazardous-capability launch decisions; that does not prove no awareness effect. |
References
- [1] Muse Spark Safety & Preparedness Report. 2026-05-14. 2026-08-24