← Evals

Evaluation profile

Google Gemini 3.8 launch — Gray Swan indirect prompt injection k=15

1sub-evals
0.546%Safety weight
0%Freedom weight
1components

Weights below are portfolio-specific global index weights.

Model score (lower is better)Predicted score

About this eval

Resistance to indirect prompt injection in agentic workflows.

Included in the behavior ranking.

Sub-evals

MeasureComponentDirection by rankingSafety weightFreedom weight
attack_success_probability_k15_pctgoogle-gemini38-gray-swan/gray-swan-ipi-k15.csv:attack_success_probability_k15_pctMeasures resistance to repeated transferred indirect prompt-injection attacks; lower values indicate fewer successful attacks.Safety: responsible_agency_control:1.000google-gemini38-gray-swanSafety: lower0.546%

attack_success_probability_k15_pct

Measures resistance to repeated transferred indirect prompt-injection attacks; lower values indicate fewer successful attacks.

RankModelValueRelative performanceProvenance
1claude-opus-54.8official
2gemini-3.8-flash5.5official
3gemini-3.8-flash-cyber6official
4claude-fable-56.5official
5claude-sonnet-56.7official
6claude-opus-4.88official
7gemini-3.7-flash9.2official
8muse-spark-1.224.2official
9gpt-5.6-sol27official
10glm-5.331.5official
11gpt-5.6-terra37.3official
12gpt-5.6-luna50official
13grok-4.651.8official
14kimi-k352.7official
15deepseek-v4-pro60.1official