NUWA Evidence record
Claude Opus 4.7
Agentic misalignment · Evidence record
Source snapshot, chart and review scope →
Original source charts and materials
Recorded measurements
- Harmful actions
- 0%[0–11]
Source denominator: 30 (unit not disclosed); 30 trials per condition
- Goal condition
- no goal
- Threat condition
- No threat
- Reasoning
- max
Inspect Evals; GPT-5-mini transcript judge · Budget not specified
Assessment provenance
GLM-5.2 Risk Evaluation Report → · External report
Actual evaluation: SaferAI · 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
0% harmful actions under no goal, No threat.
Limitations
No real people or systems were harmed; the study does not estimate prevalence in deployment.
Synthetic scenarios; 30 rollouts per condition. Intervals are transcribed from Figure 8.
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] GLM-5.2 Risk Evaluation Report. 2026-08-02. v1.1 · revised 26 August 2026 ; pp.18-19 Fig.8/§4.3