# Stealthiness Evaluation on Harmful prompts (evaluated on 3 LLMs and 4 guard LLMs)

Source: [Stealthiness Evaluation on Harmful prompts (evaluated on 3 LLMs and 4 guard LLMs)](<https://www.sota2.com/research/sota/stealthiness-evaluation-on-harmful-prompts-evaluated-on-3-llms-and-4-guard-llms>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-05-18T09:28:34.504Z

Evaluation of model stealthiness using Perplexity metrics (PPL Mean/Std) when tested against harmful prompts, specifically involving comparisons between 3 standard LLMs and 4 guard LLMs.

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

| Scope | Recorded value |
| --- | --- |
| Task | Stealthiness Evaluation |
| Dataset (including any stated split) | Harmful prompts (evaluated on 3 LLMs and 4 guard LLMs) |
| Metric | Mean Perplexity |
| Unit | SCALAR |
| Direction | Lower is better |
| Results loaded | 10 of 10 |
| Coverage | All results currently returned for this benchmark |

Evaluation settings may differ between results. The original papers and recorded configurations are needed to establish comparability. A shared metric name alone does not establish an identical protocol.

## Metrics

- [Mean Perplexity](<https://www.sota2.com/research/sota/stealthiness-evaluation-on-harmful-prompts-evaluated-on-3-llms-and-4-guard-llms.md?metric=Mean%20Perplexity>): Lower is better; unit: SCALAR.

- [PPL Std Dev](<https://www.sota2.com/research/sota/stealthiness-evaluation-on-harmful-prompts-evaluated-on-3-llms-and-4-guard-llms.md?metric=PPL%20Std%20Dev>): Lower is better; unit: SCALAR.

## Recorded results

| Method | Mean Perplexity | Result date | Report paper |
| --- | --- | --- | --- |
| ArtPrompt | 3.23 | Not recorded | 68fa751a1e209131cea65ca8 |
| Ascii | 4.13 | Not recorded | 68fa751a1e209131cea65ca8 |
| base64 | 10.46 | Not recorded | 68fa751a1e209131cea65ca8 |
| Morse Cipher | 11.81 | Not recorded | 68fa751a1e209131cea65ca8 |
| ReNeLLM | 15.56 | Not recorded | 68fa751a1e209131cea65ca8 |
| Unicode | 42.19 | Not recorded | 68fa751a1e209131cea65ca8 |
| UTF-8 | 42.19 | Not recorded | 68fa751a1e209131cea65ca8 |
| Origin | 49.9 | Not recorded | 68fa751a1e209131cea65ca8 |
| Caesar Cipher | 258.1 | Not recorded | 68fa751a1e209131cea65ca8 |
| FlipAttack | 809.67 | Not recorded | 68fa751a1e209131cea65ca8 |

### ArtPrompt

Mean Perplexity: 3.23. Result ID: 6a0adb6a47f34ce924f14ed5.

[Reporting paper](<https://www.sota2.com/research/paper/68fa751a1e209131cea65ca8.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa738f1e209131cea5d39c.md>)

Recorded evaluation configuration: Not recorded. This record does not assert independent reproduction.

### Ascii

Mean Perplexity: 4.13. Result ID: 6a0adb6a47f34ce924f14ed3.

[Reporting paper](<https://www.sota2.com/research/paper/68fa751a1e209131cea65ca8.md>)

Recorded evaluation configuration: Not recorded. This record does not assert independent reproduction.

### base64

Mean Perplexity: 10.46. Result ID: 6a0adb6a47f34ce924f14ed4.

[Reporting paper](<https://www.sota2.com/research/paper/68fa751a1e209131cea65ca8.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa70021e209131cea4887c.md>)

Recorded evaluation configuration: Not recorded. This record does not assert independent reproduction.

### Morse Cipher

Mean Perplexity: 11.81. Result ID: 6a0adb6a47f34ce924f14ed1.

[Reporting paper](<https://www.sota2.com/research/paper/68fa751a1e209131cea65ca8.md>)

Recorded evaluation configuration: Not recorded. This record does not assert independent reproduction.

### ReNeLLM

Mean Perplexity: 15.56. Result ID: 6a0adb6a47f34ce924f14ed6.

[Reporting paper](<https://www.sota2.com/research/paper/68fa751a1e209131cea65ca8.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa72d71e209131cea58d97.md>)

Recorded evaluation configuration: Not recorded. This record does not assert independent reproduction.

### Unicode

Mean Perplexity: 42.19. Result ID: 6a0adb6a47f34ce924f14ed0.

[Reporting paper](<https://www.sota2.com/research/paper/68fa751a1e209131cea65ca8.md>)

Recorded evaluation configuration: Not recorded. This record does not assert independent reproduction.

### UTF-8

Mean Perplexity: 42.19. Result ID: 6a0adb6a47f34ce924f14ed2.

[Reporting paper](<https://www.sota2.com/research/paper/68fa751a1e209131cea65ca8.md>)

Recorded evaluation configuration: Not recorded. This record does not assert independent reproduction.

### Origin

Mean Perplexity: 49.9. Result ID: 6a0adb6a47f34ce924f14ece.

[Reporting paper](<https://www.sota2.com/research/paper/68fa751a1e209131cea65ca8.md>)

Recorded evaluation configuration: Not recorded. This record does not assert independent reproduction.

### Caesar Cipher

Mean Perplexity: 258.1. Result ID: 6a0adb6a47f34ce924f14ecf.

[Reporting paper](<https://www.sota2.com/research/paper/68fa751a1e209131cea65ca8.md>)

Recorded evaluation configuration: Not recorded. This record does not assert independent reproduction.

### FlipAttack

Mean Perplexity: 809.67. Result ID: 6a0adb6a47f34ce924f14ed7.

[Reporting paper](<https://www.sota2.com/research/paper/68fa751a1e209131cea65ca8.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa751a1e209131cea65ca8.md>)

Recorded evaluation configuration: Not recorded. This record does not assert independent reproduction.

Page 1 of 1. This is an index page, not the full collection.
