Loading page…
Preferred languages
More languages supported by Google Translate
Open the selector to load the current language catalog.
Translations are provided by Google Translate.
Loading page…
Preferred languages
More languages supported by Google Translate
Open the selector to load the current language catalog.
Translations are provided by Google Translate.
external benchmark
BinaryAudit (backdoor detection in compiled binaries). Every model the catalog carries a published BinaryAudit value for, ranked by that value.
Quesma's BinaryAudit asks AI agents to find backdoors and timebombs planted in compiled, stripped binaries of real server software, using the reverse engineering tools Ghidra and radare2. The score here is the share of detection runs that found the planted code. 18 models in the ModelScale catalog have a published value.
Each point is a model with a published BinaryAudit value and complete token prices, on the 75% input / 25% output blend. Frontier models are spaced evenly along the cost axis.
18 models with a published BinaryAudit value, ordered by that value, highest first. Models the source has not scored on this benchmark are not listed — they are unmeasured, not last.
| Rank | Model | Provider | Agent with Ghidra and radare2 in a terminal, Harbor framework |
|---|---|---|---|
| 1 | Anthropic | 49.1 | |
| 1 | 49.1 | ||
| 3 | OpenAI | 45.6 |
What the benchmark measures and how its value is published, from the source’s own record.
BinaryAudit (BinaryAudit (backdoor detection in compiled binaries)) is an external benchmark published by We hid backdoors in ~40 MB binaries and asked AI + Ghidra to find them (Quesma). Quesma's BinaryAudit asks AI agents to find backdoors and timebombs planted in compiled, stripped binaries of real server software, using the reverse engineering tools Ghidra and radare2. The score here is the share of detection runs that found the planted code. Its evaluation set is described by the source as 19 backdoor and timebomb detection tasks on stripped binaries, three runs each (57 runs); 6 tooling and 8 clean-binary tasks reported alongside. The source rates its difficulty as reverse engineering without source code or symbols.
Each model's result is published by We hid backdoors in ~40 MB binaries and asked AI + Ghidra to find them (Quesma) as agent with Ghidra and radare2 in a terminal, Harbor framework. ModelScale prints that value to 1 decimal place, exactly as published, and never converts or rescales it. This page ranks the 18 models that carry a published value, highest value first; a model the source never scored is left out rather than ranked at zero. The source does not state whether a higher value is the better result, so neither does this page.
| 4 | 43.9 |
| 5 | OpenAI | 42.1 |
| 6 | Anthropic | 36.8 |
| 6 | 36.8 |
| 8 | Anthropic | 26.3 |
| 9 | Z.AI | 24.6 |
| 9 | Moonshot AI | 24.6 |
| 11 | 21.1 |
| 12 | Z.AI | 19.3 |
| 13 | OpenAI | 17.5 |
| 13 | OpenAI | 17.5 |
| 15 | DeepSeek | 12.3 |
| 16 | Anthropic | 10.5 |
| 17 | xAI | 8.8 |
| 18 | xAI | 7.0 |
Evidence key: Observed
Rows are ordered by the value We hid backdoors in ~40 MB binaries and asked AI + Ghidra to find them (Quesma) published, highest first. The source does not state whether a higher value is the better result, so this page does not either: for a benchmark that measures a rate of failure, read the table from the bottom.