Benchmarks
357 sources · 1,263 models × 3,079 benchmarks · 162,127 scoresPromptArmor's FrontierBreach measures how often indirect prompt injections hijack an agent doing enterprise work: GDPval tasks are rebuilt with the emails, files and third-party skills a deployed agent would use, and an attack planted in one of four entry points (connectors, document management systems, local files, skills) tries to make it exfiltrate data, phish, run attacker code, destroy or encrypt files, or leave a persistent foothold. Code checks environment evidence for the harmful action, a test case counts as attacked when the action happened in any of a model's five runs, and the score is the share of test cases attacked, so lower is better. Models run at medium effort; the page doesn't state the number of test cases, and attacks succeeded on 34.9% of them across the original 21 models.
No signal or difficulty estimate. This benchmark measures a behaviour or trait rather than capability, so the capability model and its validation studies leave it out.
| # | Model | Setting | Score | Source |
|---|---|---|---|---|
| 1 | Gemini 4 Carbon | medium | 2.4 | |
| 2 | Gemini 4 Argon | medium | 4.3 | |
| 3 | GPT-6 Astra | medium | 7.6 | |
| 4 | Claude Fable 5.2 | medium | 10.0 | |
| 5 | Claude Opus 5 | medium | 10.0 | |
| 6 | Muse Wonder | medium | 15.2 | |
| 7 | Gemini 3.6 Flash | medium | 22.4 | |
| 8 | Kimi K3 | medium | 31.4 | |
| 9 | GPT-5.6 Sol | medium | 33.3 | |
| 10 | Grok 4.7 | medium | 33.3 | |
| 11 | Muse Spark 1.1 | medium | 38.6 | |
| 12 | GLM-5.2 | medium | 39.0 | |
| 13 | Inkling | medium | 39.0 | |
| 14 | Beam 1.0 501B A23B | medium | 72.9 | |
| 15 | Grok 4.5 | medium | 91.4 |
Benchmarks outside its family whose model rankings track it most reliably (Spearman ρ over at least 10 shared models).