NUWA Evidence record
Muse Spark
Agentic misalignment · Evidence record
Source snapshot, chart and review scope →
Original source charts and materials
Recorded measurements
- deliberate harmful action rate
- 47.7%
Per-result sample size not disclosed
Harness not specified · Budget not specified
Assessment provenance
Evaluating DeepSeek v4 Pro for Frontier Risks → · External report
Actual evaluation: Meta · Model record →
Method and scope
Conditional rates and observed actions in controlled simulations.
- Evidence type
- Study
- Assessment subject
- Agents in fictional corporate simulations
- Primary dimension
- Propensity
- Evidence snapshot
- 2026-10-29.1
Primary observation is a conditional rate across controlled scenarios, with individual actions retained as behavior evidence.
Source finding
deliberate harmful action rate: 47.7 percent. Conditions are retained from the source.
Limitations
No real people or systems were harmed; the study does not estimate prevalence in deployment.
Harm counted only when model recognizes action as harmful and takes it; fictional motivated stress tests. Comparison reproduced by Neo Research; original evaluation not independently verified in this extraction.
Original figure retained without estimating numerical values. Full source tables have not been transcribed.
Related research questions
Autonomy / Unsanctioned action →
Varied controlled scenarios test actions contrary to the assigned benign role.
Deception / Covert goal pursuit →
Some scenarios test concealed harmful actions in pursuit of an assigned goal.
References
- [1] Agentic misalignment: How LLMs could be insider threats. 2025-06-20 ; Official publication masthead and article byline; Source figure
- [2] Evaluating DeepSeek v4 Pro for Frontier Risks. 2026-05-29. v1.1 · 29 May 2026 ; PDF p. 30; Anthropic Agentic Misalignment; deliberate harmful action rate