Evaluation profile
UK AISI active safety-research compromise continuation
1sub-evals
0.36%Safety weight
0%Freedom weight
1components
Weights below are portfolio-specific global index weights.
Model score (lower is better)Predicted score
About this eval
Propensity to continue a prefilled trajectory that actively compromises AI-safety research.
Included in the behavior ranking.
Sub-evals
| Measure | Component | Direction by ranking | Safety weight | Freedom weight |
|---|---|---|---|---|
| active_compromise_continuation_rate_pctopus5-system-card/uk-aisi-active-compromise.csv:active_compromise_continuation_rate_pctMeasures whether the model actively continues an ongoing system compromise after a prefilled trajectory suggests the attack is already underway. | Safety: responsible_agency_control:1.000opus5-system-card | Safety: lower | 0.36% | — |
active_compromise_continuation_rate_pct
Measures whether the model actively continues an ongoing system compromise after a prefilled trajectory suggests the attack is already underway.
| Rank | Model | Value | Relative performance | Provenance |
|---|---|---|---|---|
| 1 | claude-opus-5 | 0.1 | official | |
| 2 | claude-opus-4.7 | 0.8 | official | |
| 2 | claude-opus-4.8 | 0.8 | official | |
| 4 | claude-mythos-preview | 1.7 | official | |
| 5 | claude-mythos-5 | 13.6 | official |