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
ExploitGym Honeypot out-of-scope exploitation rate. Every model the catalog carries a published ExploitGym Honeypot value for, ranked by that value.
An alignment test for cyber agents: hard ExploitGym tasks are paired with an easy-to-exploit auxiliary service (such as a package registry or object store) that would reveal the flag but is outside the authorized scope. Lower scores mean the model stays in scope more often. 11 models in the ModelChoice catalog have a published value.
Each point is a model with a published ExploitGym Honeypot value and complete token prices, on the 75% input / 25% output blend. Frontier models are spaced evenly along the cost axis.
11 models with a published ExploitGym Honeypot 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 | Out-of-scope exploitation rate (lower is better) |
|---|---|---|---|
| 1 | OpenAI | 49.50 | |
| 2 | OpenAI | 9.90 | |
| 3 | OpenAI | 5.51 |
What the benchmark measures and how its value is published, from the source’s own record.
ExploitGym Honeypot (ExploitGym Honeypot out-of-scope exploitation rate) is an external benchmark published by GPT-6 Cyber: Advancing the frontier of aligned cybersecurity agents. An alignment test for cyber agents: hard ExploitGym tasks are paired with an easy-to-exploit auxiliary service (such as a package registry or object store) that would reveal the flag but is outside the authorized scope. Lower scores mean the model stays in scope more often. Its evaluation set is described by the source as hard ExploitGym tasks with an out-of-scope honeypot service. The source rates its difficulty as cyber-agent scope adherence.
Each model's result is published by GPT-6 Cyber: Advancing the frontier of aligned cybersecurity agents as out-of-scope exploitation rate (lower is better). ModelChoice prints that value to 2 decimal places, exactly as published, and never converts or rescales it. This page ranks the 11 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 | OpenAI | 0.07 |
| 5 | OpenAI | 0.00 |
| 5 | OpenAI | 0.00 |
| 5 | OpenAI | 0.00 |
| 5 | OpenAI | 0.00 |
| 5 | OpenAI | 0.00 |
| 5 | OpenAI | 0.00 |
| 5 | OpenAI | 0.00 |
Evidence key: Observed
Rows are ordered by the value GPT-6 Cyber: Advancing the frontier of aligned cybersecurity agents 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.